AI Agents Full Course 2026 | AI Agents Tutorial for Beginners | How to Build AI Agents | Simplilearn

Key Takeaways

Hey everyone, welcome to our agent school course by simpler. Are you curious about AI agents and how they're changing the world right now? Then this video is just for you. AI agents are everywhere and in huge demand. Learning how they work could open doors to amazing jobs in AI, automation, and tech. In this full course, we will start with the basics. what AI agents are, how they interact with the surroundings, and how large language models work, power them behind the scenes. We'll also dive into what generative AI agents really do, cover the full road map, and even go over some math basics like neural that will help you understand about machine learning. We've also included tutorials on key machine learning like GANs, NLP, K, and confusion matrix explained in easy way. But before we commence, if you're interested to master the future of technology, then the professional certificate course in generative AI and machine learning is the perfect opportunity for you. Offered in collaboration with the NICT Academy, IT Kur, this 11month live and interactive program provides hands-on expertise in cutting edge areas like genative AI, machine learning tools like chat, GPT, D2, and even hugging face. You gain practical experience through 15 plus life projects, integrated labs and master classes delivered by esteemed IT Kpool faculty. So hurry up and enroll now and find the course link in the description box below. Learns YouTube channel. In this tutorial, we will learn about AI agents. Before that, let's take a quick look at what exactly they are. As AI continues to evolve, there are no longer just chatbots responding to your commands. It's about intelligent systems that can understand objectives, make decisions, and take actions. AI agents are autonomous systems that perceive their environment, process information, and take actions to achieve specific goals. They use machine learning, natural language processing, and reasoning techniques to make decisions. AI agents can be responding to stimuli. They are widely used in virtual assistance, automation, and decision-m systems. Advanced AI agents like conversational AI and autonomous agents can adapt and learn from interactions. That said, if these are the type of videos you'd like to watch, then hit that like and subscribe buttons and the bell icon to get notified. Also, for your information, if you want to upskill yourself, master generative AI and artificial intelligence skills and land your dream job or grow in your career, then you must explore Simply Learn's code of various generative AI courses and certifications. Simply learn offers various certification programs in collaboration with some of the world's leading universities like Kurdu IIT Kpur, Guati and many more. Through our courses, you will gain knowledge and work ready expertise in skills like advanced Python, machine learning, generative AI and over a dozen others. That's not all. You also get the opportunity to work on multiple projects, learn from industry experts working in top tier data and product companies and also academicans from top universities. After completing these courses, thousands of learners have transition into an AI and machine learning role as a fresher or moved onto a higher paying job and profile. If you are passionate about making your career in this field, then make sure to check out the link in the pin comment and description box below to find a generative AI program that fits your experience and areas of interest. Now, without further delay, let's get started. Firstly, let's have a brief discussion about the agenda for today's session. So, these are the following points that we will be discussing in the AI tutorial for beginners. Firstly, we will get started by understanding what exactly is an AI agent. Next we will learn the critical differences between AI agents and large language models. Followed by that we will understand how exactly an AI agent works. Then we will have a brief overview on the future and risks of AI agents. Followed by that we will have some final thoughts on AI agents and then we will switch to a live hands-on demonstration on how to build AI agents. I hope I made myself clear with the agenda. Now let's get started. Firstly, what exactly is an AI agent? So, at its core, an AI agent is a software program designed to perform tasks autonomously. Unlike traditional software that strictly follows predefined rules, AI agents can make decisions based on understanding, interaction, and real-time data. They leverage cuttingedge AI models like GPD, Claude, Gemini to analyze, process and execute tasks dynamically. Think of it as a digital assistant that doesn't just follow rigid instructions like try saying check if Tony is available on Friday and book a meeting. This is a usual approach, right? Instead of following a routine approach where you try to say to your digital assistant this way, check if Tony is available on Friday and book a meeting. This is a routine approach where you say to your digital assistant to check the availability of Tony and book a meeting which is quite straightforward and simple to understand. But in case if you take it to a much complex version. But if you just change a couple of terms and make it a little more complex like try saying schedule a meeting with Tony at the earliest convenient time in the next month. Right? It might seem a little similar to the previous one but here the AI agent will have to follow some steps like this one. Check the calendar for available slots. The AI agent will log into the recipient's account which is Tony's account and checks if there are available slots for the meeting. Then it will determine the ideal time followed by that it will send out invitation which happens all autonomously. In the previous case directly the digital assistant will check for a Friday in the calendar and books a meeting for the time that you have provided regardless if availability is there or not. This is the difference between a digital assistant and an AI agent. Now let's move ahead into the next topic of today's session which is about how are AI agents different from LLMs. While AI agents use large language models like GPD4, they go beyond simple text generation. Here are the key differences. LLMs generate responses based on trained data but do not actively interact with the real world. Whereas on the other hand, AI agents can access external data, plan tasks, and execute actions all in real time. This is one of the critical defenses between LLMs and AI agents. Now, for example, if you ask LLM about a sports match that happened yesterday, it may not have an updated knowledge. Well, on the other hand, AI agents can fetch live data from the web, databases or APIs to provide an accurate answer. This is how they behave and perform. Also, there is something called hallucinations. You might be wondering if this term is related to humans. Yes, it is related to humans. But for a change, LLMs can also hallucinate. For example, if you ask unrealistic question, let's say there are different types of apples, Fuji, Macintosh, Black Oxford, right? If I write, tell me about watermelon apple or orange mango which are unrealistic or in a way different fruits, chances are there where an LLM can give you some false or unrealistic information that may sound believable. This is termed as hallucination. And of course, recent LLMs have a scale to measure their hallucinations as well. Now, let's shift to the next point of today's discussion, which is about how do AI agents work. Now, AI agents function as a sophisticated problem solvers with four essential capabilities. Let's start with planning, then interacting with tools, then memory and knowledge access, and lastly, executing the actions. Now, let's go through each one of these steps in a detailed way. Firstly, planning. AI agents begin with a goal. Whether it's researching market trends or drafting an email to break the goal into actionable steps following a chain of thought process to optimize the execution. This removes the need for predefined human triggers. The next stage is interacting with the tools. Unlike basic AI models, AI agents can interact with tools, browse the internet, query databases, and use APIs. This extends their capabilities beyond static knowledge. Now the third one is about memory and knowledge access. AI agents can retrieve and store information using techniques like R A G or also known as the retrieval augmented generation. This allows them to integrate company specific data market research or customer support queries to generate up-to-date responses. Now the last and important step, execution of actions. Beyond just processing information, AI agents can take actions, writing reports, managing applications, or even communicating with other AI agents. This makes them powerful automation tools capable of independently handling complex workflows. Now, let's move to the next part, which is about the future and risks of using AI agents. AI agents represent a new era in automation and efficiency. However, their autonomous nature comes with risks. Without proper safeguards, an AI agent could misinterpret objectives. For example, if instructed to achieve world peace, an unregulated AI might arrive at extreme solutions. Ensuring human oversight, ethical programming, and responsible deployment will be crucial to leveraging AI agents safely and efficiently. Now let's go through the next point where we will have some final thoughts before getting started with the live demonstration. So coming to the final thoughts, AI agents are redefining how we work and interact with technology where their ability to plan, interact with tools, store knowledge and execute tasks. They are far more advanced than traditional LLM. Now that we have discussed the theory part of today's session, let's move ahead to the next part where we will have some hands-on demonstration on how to build AI agents. So today we will be building an AI agent that can help you browse through your internet. Now we will be using browser use because of its highest accuracy based upon its competitors. So it has about 89% compared to other competitors. Right? So there are three different plans open source, pro and enterprise. So for now we will be going with the open source version which is readily available on GitHub. So here you'll have the complete repository. So this is the browser use. Here we have an official page of browser use where we have the connection to quick start and all the steps are arranged in learning order. You can just follow the steps and get ready with the environment for building your AI agent. So we have opened up the command prompt. First we need to check Python if it is available or not. So it's one of the prerequisites. So Python version we have about 3.9. So uh according to the prerequisites we need to have anything above 3.10. Not a problem. Let's download Python. So here just go to Windows. I'm working on Windows. Windows 3.13. It should take a couple of moments. And the playright is also one of the prerequisites and UV is also one of the prerequisites. It's a Python package which helps us to create a virtual environment. So if you just follow these steps, you can also create a virtual environment for yourself to run your AI agents. Everything is just follow the steps, right? Let's have a quick look on where exactly we are on the Python installation part. There you go. Python has been successfully downloaded. You just have to run the execution file. And make sure you click on that add python.exe to path so that you don't have to manually create the uh you know environment variable for that. It'll automatically create and everything is as simple as that. Just set up follow the default methods of setting up Python in your local Windows operating system and it should be done. Just give it some time. There you go. We are shortly closing the first step which is adding Python to your Windows operating system with the latest version. There you go. The installation was successful. Now let's start command prompt and check if we have successfully updated Python version or not. Simple the same command Python version. Enter. There you go. So we have 3.13. Now rest of the part will happen on the command prompt. So let's expand it and also have the page open for browser use so that we can proceed with steps and install everything. So we'll go to the GitHub page. So before going to the GitHub page, let's go to the official web page where we need to follow some steps to prepare the environment first. So here we have uh a command to set up an environment for UV. Okay, UV is not internal. So let's try to go through uh Stack Overflow to resolve this issue. I think uh we might have to install UV first or let's give it another try. Let's run the same command. Okay, the same issue. So I think we have to update uh UV first. So let's copy paste uh that particular code from UV website. I'll show the So I think we have to set the UV environment first. So we have an official documentation here which shows some steps to be followed to set up UV. So just copy the code from the official website and run these on the command prompt as mentioned here. Just enter. So the first package has been successfully downloaded and yeah everything is installed. Now let's go back to the official web page and try to execute the second command in the row. So that's the one request a specific version by including the URL. So for a specific version we need to copy it and then we have the pip install UV and lastly we can proceed with the uh you know uh environment here Python. So here's the uh next one. I think we might have copied the wrong code. So, let's copy the right code here. Paste it there. Execute the second code as planned. Downloading. It might take a little time for unboxing. Expansion in progress. There you go. Everything is installed. Now, we can get started with setting up the environment for Python and AI agents. Go to the quick start for browser use. Oh, we missed one thing. We need to pip install UV first. Yeah, there you go. Okay, I think we are facing an issue again. Let's go with uh the Windows version. I think this should work. Yeah, collecting and downloading the metadata. It might take a little while. Once we install UV, then we can continue with V, which is virtual environment for Python and our AI agent. uh we are not having that in here. Let's go to the browser use official page. So yeah, I guess we are all set. I think uh we can get started with executing the virtual environment. So let's go to the quick start web page. So copy the code. Just type it down there and hit enter. That's the one. So we have 3.13. Uh will it execute? Let's give it a try though. Yeah, it is executing. I think it took a little time. There you go. Now, we need to activate the virtual environment. For that, you can see activate with when these scripts activate, right? You just need to copy paste that code. Don't forget the dot. It's not copying. Don't worry. We can just manually type it down if it doesn't copy. Or you can also go to the web page. Actually, it is also mentioned in the steps. You can just copy it from the web page and paste it here or you can choose to type it down. So we have successfully in the virtual environment. So you can see the username is present inside the brackets. So that means you are in the virtual environment. Now you can proceed with the further steps of setting up the environment before you build an AI agent. Now we are back on the browser use. So here that was the step that we need to take. Now copy the browser use code which is mentioned here to download the dependencies. Enter. So packages are getting prepared to unbox. I think shortly the download will be finished for browser use. There you go. We're good to go now. Now let's proceed with the next command. So the next command is about installing playrite. Just copy the code and the next part will be about creating your AI agenda. So playright is the last dependency that you need to install. Again it might take a little while. Let's wait. There you go. Playright has been successfully installed. Now the last and final step is getting started with building the AI agent. Now talking about the AI agents, you can go to browser use and go to some examples from the repository or you can choose to copy this code from browser use official documentation and customize your AI agent. So here if we go back to browser use repository go to examples here you already have a lot of examples or AI agents built by browser use. So you can click into use cases and you will find a lot of them. So identifying capture, shopping or you know checking out Wikipedia pages for information, scrolling and searching using the search bar to search for a specific keyword, right? All of those are readily available in browse website or repository. You can give it a try and learn things from that. So you can do this by cloning the repository. For that just click on the code and here copy this code extension and run it on your window. It is running in virtual environment. So it can be cloned by using get space clone space the command. Okay. Git is not recognized as an internal or external command. I think we are supposed to install git as well. For that let's try to open a new command prompt and try to install git in that. So there you go. We have a complete detailed procedure to install and run git in the Windows operating system. So these are the steps to be followed. Let's go to get official website and download git first. So I'll go with the 64-bit version for my system. Yeah, let's download it. So, basic installation procedures following the default steps. There you go. Finish. Now, go to a new Windows command prompt and check for the version. Okay, it's not still recognized because I think we might have to set up the environment variables for this. Unlike Python, Git doesn't provide an automatic path. So, you can do that by selecting the environment variables. Not for the account. Yeah, not for the account. We might have to go for the system variables here. Let's go back. Yeah, system variables. Create new. You can add the variable name as okay. Let's cross check one last time. Cancel. Yeah, you had to go into system variables and create new. paste the path from git right so for that we might want to go into the program folders and then get folder and cmd that's the location I'll just name name it as git and this is the file path let's use this lowest case okay done now let's close the earlier command prompt and uh yeah it's not working so let's close and Restart the command prompt as provided here where if I close and reopen your command prompt or PowerShell. Okay, let me start a new command prompt. This is a virtual environment. Let's not mess with it. So, get -y version. So, we have successfully installed git. And now let's try to clone the repository in the virtual environment. I think we might have to restart the Yes, there is a requirement where you have to restart the virtual environment as well. Not a big problem. Let's try to do that. So before we restart uh the virtual environment, let's open a new command prompt and try to clone it there and then replicate the same on the virtual environment. So this is the new uh command prompt. Let's copy the code and we will use the get clone keyword. So it is getting cloned. So it's working. So now we can retry this on the virtual uh environment. So I'll close the entire command prompts. So there you go. In the virtual environment, the file already exists. So we have successfully cloned that file. Now we will go back to desktop and create a new folder empty folder and rename it as AI agent. Right? So far we have used and uh set up all the environments required by browser use. Now to create our AI agent we also need an interface for that browser use I have already provided a web user interface web UI that's the second link right. So we will be downloading and installing that into this particular file. So uh let's start the command prompt again and go back to desktop. So you need to just provide cd desktop. So this will change the directory to desktop and here on the desktop provide the location of AI agent. So I have user AI agent. You will be inside this folder now. Yes, there you go. Now you will need to execute some commands where you will install web UI or browser UI into this folder. Go to the homepage browser use web UI. This is the link GitHub link and here the same procedure you will be provided with stepbystep guidance to install web UI and you'll be ready as soon as you do it. So we'll be cloning it. So we have already installed uh get clone. So get clone space code run to clone all the repository into this AI agent folder. There you go. It's done. For cross verification, you can open the web UI folder in desktop and you'll have all the dependencies. Now let's continue with the next steps. So in the next step, we will change directory to the web UI folder. We'll be inside the web UI folder now. There you go. So whatever the next steps that we will be executing will be inside the web UI folder. So they say we have to enter into the virtual environment again. So UV. So I think uh they recommend having UV. So we have already installed UV. But anyways for safer side let's try to execute this command prompt. So we have two options Mac Linux Windows. So we'll go with Windows. Select the code. Paste it here. Done. It should take a couple of moments. So as the second procedure, let's copy the second code segment as well. And uh let's go back to the command prompt and see what's the status. And there you can see already everything has been installed. Let's also do the second step as well. So we've pasted the second code segment and it's unpacking. So let's paste the third snippet which is about creating the virtual environment. So for that we will be installing UV as we have done previously. Yeah. So let's go back to the web UI interface and u copy the steps provided here. So we will be creating a virtual environment using this particular code snippet which is UV when v Python 3. Erase the previous code. Enter. and we will be shortly inside the virtual environment. Let's activate when V. So if you're going through Windows, you just have to select scripts. And if you're using Linux and Mac OS, there's a different statement. So we'll be using with scripts. There you go. We are inside a virtual environment. And here the username is web UI, right? So we're going to the next steps. We need to install some dependencies. For that we need to use uip install our requirements txt. Just copy paste the code. There you go. The dependencies are getting downloaded. It might take a little while. There you go. Uh the dependencies might get downloaded and unpacked shortly. Just give it a little time. I guess it's done. Now let's go with the next procedure where we will have to install playright once again. Go back to the uh GitHub page. So the code has been copied to install playright. And next we will have uh Docker. I think we'll go with local setup. Right? So if you are transferring the AI agent to a different user and making the agent public then you can go with the docker application. But if you're creating an AI agent for yourself and your purpose only then you can go with a local setup where you can go with the local setup and it will be active only in your PC in local uh environment. So uh this will be your port play right has been successfully installed I guess. So uh here you are creating a user interface for yourself and once you execute this command the command prompt will give you an IP address. So it will be in the form of uh local host and you just have to copy paste that particular URL in your browser. Just wait for a little while while it gives you the URL. There you can see running local on URL. So you need to copy this URL http colon/12700.1 double quotes 78. That's your port number. Copy that and enter on any of your local browsers. So I'll open a new uh window over the browser and I'll paste it over here. There you go. Now you can see a web UI is successfully running with an interface for your agent. Now your agent is not completely done yet. You can see you have some maximum run steps and maximum actions per step. If you're providing a command, it it runs a certain number of steps, right? You can can customize. You can either increase or decrease, right? Those are a few things you can choose to customize. So I'll take the maximum actions per second to 15. Now you can see the LLM settings. This is the most important part. By default we have OpenAI here. So I'd like to go with Google and inside that we have the model 2.0 flash. And uh now we need something called as an API key which can be created. So uh you just have to go to Google studio for that. Let's open a new tab, Google Studio. Select the first link. You might have to sign in here. So, don't worry about the pricing. This is open source, free of cost. Let's quickly sign in. So, now I've successfully signed in. So, here you can just create uh get API key. So, you're using Google AI Studio here. So we'll try Gemini and get API key. Can close this prompt here and accept all the terms and conditions. Now just click on get API. Scroll down create API and now you will have your API created successfully. Let me remind you this API will be only for your personal use. So don't share it with anyone. Just copy your API key and you can see even uh the windows says don't share with anyone. So copy the API key and go back to your browser and paste the API key right there. Now you can choose to run your agent here. You can type down your steps to follow for your agent. So let me type in go to google.com and search for simplylearn.com or simply learn website. So if you go to the command prompt here you can see stepby-step execution. So it has started a new uh window for Google browser and you can see it is in signed out mode. Clearly it's accessing the search prompt and I'm not typing anything. It's doing everything by itself. Simply.com. Okay. It's selecting the enter button or search button by itself. There you go. It has identified the search results. And now it should identify the page as well which I think it is already in progress. Yeah, it has already started the Simply Learns homepage. There you go. And you can also find a course that you are looking for as well. Great. Now let's take a step further. Let's stop this application here and let's try to uh you know take a step further and try to provide an email address and u even the login credentials and ask the AI agent to use that particular credential login and send out an email. Since we have already seen this particular step has been executed in signed out mode, right? So let us provide a command for our AI agent with the credentials in such a way that it can open a new window and can log into the Google Gmail account and send out an email. Right? Let's try that. So I'll erase the previous prompt. Go to google.com. Enter step two. Log to Gmail. And let's provide the email address along with the password. Now let's state compose an email to uh let's also provide a random email address here as well. subject will be enrolled to AI and ML program. Let's try to run the agent now. I see it has opened a new window. But uh okay, let's check if it selects the sign in option. Yeah, it is selecting the sign in option. It has recognized it. Okay, it opened the sign in page as well. It is trying to type in if I'm not wrong due to some or the other security restrictions it should not be able to perform it but still I can see uh the cursor highlighting in the window. Let's not lose hope. Let's see what it can do. No, it stopped automatically. So that means u it does have some limitations but it's anyways good. So far it's so good. So in case if you wanted to uh you know kind of book some credentials or book some items on the online website or search for the best uh items with low cost some uh preverified commands which are not security concerned then so far so good this particular agent is very good for you and if you are taking a step beyond where it's ethically not logical to perform then the agent is choosing to withdraw itself from that command. Picture this. You've got a virtual assistant. Let's call it your smart AI buddy. So, this assistant doesn't just follow basic commands like setting reminders or sending emails. This AI can think for itself, make decisions based on the situation, and even learn from every interaction it has. It can predict what you might need, solve complex problems, and even adapt to new challenges without any input from you. That's the magic of agentic AI and it's not thing of the future, it's happening right now and the exciting part is you can build it too. So in today's video we will be taking a deep dive into the road map to learn agentic AI step by step. So here's what we will cover. First of all, we'll start with the basics of programming focusing on the most beginnerfriendly language which is Python. Then we look into the machine learning and natural language processing to teach your AI to learn and understand human language. The next step is deep learning and transformers, the powerful models behind most modern AI systems. Then we'll get hands-on with generative AI where we will learn how to build AI that creates text, images, and many more. We'll also see how agentic AI works, how you can create intelligent agents that can interact with the world and make decisions on their own. and we'll wrap it up on how to deploy your AI agents on the cloud to make them live and ready for real world use. So by the end of this video, you'll exactly know how to create and deploy your own AI agents. So let's dive in and start this exciting journey together. So let's start with the very first step which is learning Python. Now before we can build intelligent AI agents, we need to learn how to communicate with our computer and that's where Python is the language we will use to do that. Python is one of the most easiest programming languages to learn and it is widely used in the AI development. It's very simple, easy to read and most importantly it has got huge community of developers which makes learning and troubleshooting much more easier. Python also allows you to focus on problem solving rather than worrying about the complex syntax. But what exactly is programming? At its core, programming is just writing a set of instructions to tell a computer what to do. Think of it like teaching your robot how to follow a recipe. You provide the step and it executes them. Now, for agentic AI, you won't just need to know the basic Python. You would need to get comfortable with some specific libraries and concepts that are most relevant to AI development. Now these would include libraries from data manipulation, machine learning and natural language processing NLP. So let's dive into some of these. Now we'll just check the libraries and use cases relevant to agentic AI. So the very first library you will encounter is NumPy which will help you to handle and perform calculations on large data sets. Agentic AI often deals with massive amounts of data and Numpai is great for making these calculations faster and much more easier. Next, we've got pandas. Think of pandas as a superpowered Excel for Python. So, it will help you to organize, clean, and even analyze data very efficiently. Now, since agentic AI requires handling lots of data to make decisions and learn from its environment, pandas is a must know. Then we've got Matt plot lib and seabbond for data visualization. Now while you might not immediately think of graph when building an AI agent, understanding data trends visually is a huge part of training and refining your model. Now for actual AI development, the key libraries are scikitlearn and tensorflow. Now, scikitlearn is great for machine learning projects where your AI needs to learn patterns from data. TensorFlow, on the other hand, is a powerful tool for building neural networks, which is exactly what you will use for deep learning models like those that power agentic AI. Finally, we've got NLTK and Spacey, which are popular libraries for natural language processing. Now since Asgentic AI needs to interact with humans and understand language, these libraries also help your AI process text, understand sentences, and even generate humanlike responses. Now, where should you start learning Python? Now, if you are a complete beginner, Simplearn offers a great Python for beginners course that walks you through the basics and builds up the more advanced topics including machine learning and many more. So, if you want a link, I'll just mention the link in the description box below. So you can just directly check it. Now let's move on and talk about some key Python concepts for agentic AI. Now to get started with agentic AI, here are the key parts of Python you will need to focus on. First we've got variables and data types. You would need to understand how to store and manipulate data. Then we've got loops and conditionals. These are very important where you will need your AI to make decisions on repeat task just like a robot performing a task continuously. Then we've got functions and classes. Now functions will help break down task where classes will allow you define your own AI object and behaviors like defining an intelligent agent. Then we've got libraries and modules. So Python's power come from its libraries. You would need to learn how to import and use them effectively. For agentic AI, you need to focus heavily on understanding data structures such as list and dictionaries because you will be processing a lot of data. You also need to get comfortable with functions classes to organize your AI behavior and the abilities into neat reusable pieces. So now that you know how to communicate with computer using Python, it's time to teach our AI agent two major technologies which is machine learning and natural language processing. So, what exactly is machine learning? Machine learning is a type of artificial intelligence that allows computers to learn patterns from data and improve over time, all without having to be told every single step. Think of it like this. Imagine you want to teach a robot how to recognize if a review is positive or negative. Instead of telling it exactly what to look for, you will give it a lot of example positive and negative reviews. and then it starts to learn from them. The more examples you give, the better it gets at recognizing them on its own. That's what called machine learning. And here's how it works in Python. Let's say we want a robot to learn whether a review is positive or negative. So based on some example reviews, we feed it to the data and let the model figure out the rest. It's just like teaching it to read between the lines and identify emotions in the text. And this is how exactly agentic AI works. So agentic AI will use the machine learning to learn from the data and make decisions whether it's making recommendation, answering a question or even automating task. Machine learning is at the heart of it all. By teaching the AI to recognize pattern, it becomes very important and smarter over time better at making decisions on its own. Next, let's talk about the natural language processing. Now that our AI agent can understand and interact with us in human language, NLP is the technology that allows computers to read, understand, and even generate text just like a person. Think of it as your Siri or Alexa that can understand what you're seeing and respond back to you. That's NLP in action. For example, let's say you ask Alexa, "What's the weather like today?" So the NLP will help Alexa understand that you're asking for weather information and then it responds accordingly. Now with the help of agentic AI, NLP is very important because AI needs to understand conversation commands and questions from humans in natural everyday language. So whether it's answering customer queries,uling appointments or even writing emails, NLP allows your AI agents to interact with the world in a very meaningful way. For example, you can use these libraries to build an AI agent that can read a customer review and tell whether it's positive or negative, just like we did in our machine learning example. Or you can even create an AI that responds to customer inquiries by understanding their questions and generating appropriate responses. Now that we have covered the basics of machine learning, let's take things up a notch with deep learning, a more advanced form of machine learning. So deep learning powers some of the most impressive AI applications you see today just like self-driving cars, virtual assistant like Siri and Alexa and even AI generated text. At the heart of deep learning is something called neural networks. Think of neural networks as a mini brain for your AI. They're designed to process data in a way that mimics how the human brain works. What makes them so powerful is that they have multiple layers that allow them to recognize patterns in data much more accurately than traditional machine learning models. Here's an example. Imagine you are teaching an AI to recognize whether a picture is of a cat or a dog. Now, with the help of neural network, the AI doesn't just look at the whole picture at once. It breaks down the picture into smaller parts starting from simple features like the edges and the colors and then building up to more complex features like the ears or fur. Each layer of the neural networks adds another level of the understanding and that's how deep learning work and that's why it's so good at task like recognizing images, translating languages and even generating creative content. Let's say we have an AI that can identify objects in images like self-driving car. When the car's AI see an obstacle, it needs to quickly recognize whether it's a pedestrian, a vehicle, or something else. Now, using deep learning, the AI can process this information accurately and in real time, allowing it to react quickly and prevent accidents. But wait, there's more. Let's talk about transformers. A specific way of deep learning model that has completely revolutionized the world of AI. Transformers are gamechanging because they're incredibly good at handling sequences of data like text or speech. For example, language translation or text generation. These are the things transformers excel at. So transformers are behind some of the most powerful AI models you have probably heard of like GPT3, the model behind chart GBT and Google Bird. These models can understand and generate humanlike text with remarkable accuracy. Now let's break it down for you. Imagine you want to create a chatbot that can have a real conversation. The challenge is that conversations don't happen in a straight line. People talk in paragraph, ask questions, reference past comments. So there's a lot of context to keep track of. Transformer solves this problem by looking at the entire sequence of words, not just one word at a time and understanding the relationship between them. This allows the model to grasp the full context of a conversation, just like how humans communicate. Here's an example. If you were to give a transformer a prompt like once upon a time there was a young prince, it could generate the rest of the story in a coherent engaging way just like a human author might do. Now let's talk about transformers in agentic AI. Let's connect all of this to agentic AI. Now since agentic AI is all about creating intelligent agents that can interact with humans, transformers play a huge role in making these interactions natural and meaningful. So whether it's like having a chat with a virtual assistant or having an AI solve a problem through text, transformers are what allows these agents to understand complex language and generate responses that make sense. So after you have understood the deep learning and transformer models, let's move on. And at this point, we are ready to get hands-on with generative AI, a type of AI that creates something new like text, images, or even music. So in this part we will explore how AI generates humanlike text, images and even music all on its own. Frameworks like lchain and langraph are very important and perfect for building genai models. Now let's talk about something really exciting which is agentic AI. So this is where things get truly magical because agentic AI is a special type of AI that goes beyond just following instructions. So you would need to understand what exactly is agentic AI and how AI is different from the other AI systems that has the ability to make decisions, interact with the environment and learn on its own. For example, an agentic AI could be used to schedule appointments. It could look at your calendar, check availability and automatically book a meeting without needing to be told every single time. Now, agentic AI can even manage task across different platforms. Imagine it handling emails, setting reminders, and organizing your work all at once. It's a really powerful tool for businesses and individuals looking to save time and automate task. Now, to build agentic AI, you will need to use frameworks like feed data, crew AI, and even language. Now, these frameworks make it easier to create AI agents that can think, learn, and make decisions on its own. So once you've built your AI agent, the next step is to deploy it in the cloud so that it can scale and work on the user globally. Cloud platforms like AWS, which is Amazon Web Services, and Google Cloud provide powerful tools to host and run your AI models in a way that ensures they can handle large amounts of data and even traffic. So when you deploy an AI agent on a cloud, it's not just running on one computer. It's running on powerful servers that can handle multiple users at the same time globally. And why is it important? Imagine you have built an AI agent to handle customer service question. Now, if you only want to run it on your personal computer, it won't be able to handle lots of customers all at once. But when you deploy it on the cloud, you can make sure it scales up automatically, handling thousands or even millions of users without crashing or slowing down. Most of you are already familiar with Alexa, Siri, and Google Assistant. Some of you are used to smart devices like an air conditioning system that can sense your absence in the room and automatically turn it off. Have you ever wondered who these are or how do they do all these and how can they be so human? Don't worry, this video will help you learn everything you need to know about these artificially intelligent devices also known as AI agent. Agent is a freestanding program or entity that interacts with its surroundings by sensing them with sensors acting with actuators orectors and so on. Coming to software agents, this agent acts on sensory inputs such as file contents, key strokes and network packets I repeat and network packets it has received by acting on those inputs and showing the results on a screen. Coming to human agent, we are all agents of codes. Humans contain sensors like their eyes, hearing and other organs as well as actuators like their hands, legs, mouths and other bodily parts. And third one is robotic agents. Robotic agent feature a variety of highquality motors that serve as actuators as well as cameras and infrared range finders that serve as sensors. This is all about examples of AI agents. Agents must to certain rules. Let us list them out. Rule number one, an AI agent needs to have an environmental perception. Rule number two, decisions must be based on observations of the environment. Rule number three, action should follow decisions. And coming to rule number four, the AI agent's action have to be logical. The best possible outcome comes from making rational decisions that maximize performance. Now that we all know the rules to follow, we will understand types of agents. First is learning agents. A learning agent learns from its past experience. It is used in the gaming industry as this industry needs the most reliable and tested agent. Next is reflex agent. These agent focus on the now and forget the past. They answer by applying the evident condition action principle. When a user starts an event, the agent refers to a set of pre-programmed criteria and rules leading to pre-programmed result. You might have heard about tic tac to and the decision regarding that based on reflexive agent model based agent. These agents select their behaviors in a similar manner as reflex agent but they have a more thorough understanding of their surroundings. The internal system has an environmental model that integrates with the agent's past. Next is goal-based agent. These agent expand on the data that a model based agent maintains by adding goal information or data describing desired outcomes and circumstances. Next is utility based agent. They are similar to goal based agents but they also provide an additional utility metric. This evaluation ranks each potential consequence in relation to the desired outcome and chooses the curse of action that optimizes the result. Examples of rating criteria includes factors like success probability or the quantity of resources needed. Now we know the types of agents. It's time to know its function. So let us learn about structure of an agent. Many agents uses the PA model in the structure. P is an acronym for performance measure environment actuators and sensors. For instance, let us take vacuum cleaner as an example. Performance, cleanliness and efficiency can be considered as performance. Environment rug, hardwood floor, living room, these can be considered as environment for vacuum cleaner. Actuator brushes, wheel and vacuum bag in a vacuum cleaner can be considered as actuators. Sensors, dirt detect sensor, bump sensor. Next we will see how to improve the performance. We only need to question ourselves. How do we increase our performance in a task? When confronting the problem of how to improve intelligent agent performances. Clearly the solution is straightforward. We carry out the task, recall the outcomes and make adjustments in light of our memories of prior price. The same manner that humans improve, so do AI agents. By saving its previous responses and attempts, the agent improves and learns how to react more effectively in the future. Artificial intelligence and machine learning conversed how machine learning can now understand and generate humanlike text, you are in the right place. From chatboards like chat jeepy to AI assistant that powers search engines, LLMs are transforming how we interact with technology. One of the most exciting advancement in this space is Google's Gemini or OpenAI charging large language model designed to push the boundaries of what AI can achieve. In this video, we will explore what LLMs are, how they work, and why models like Geminy are critical for the future of AI. Google Gemini is part of a new wave of AI models that are smarter, faster, and more efficient. It is designed to understand context better, offer more accurate responses, and integrate deeply into service like Google search and Google Assistant, providing more humanlike interactions. So, we will break down the science behind LLMs, including their massive training data set, transformer architecture, and how models like Gemini use deep learning innovation to change industries. Plus, we will compare Google Gemini to other popular LMS such as OpenAI Chat GB models, showing how each of these technologies is used to power chat bots, virtual assistants, and other AIdriven application. By end of this video, you will have a clear understanding of how large language models like Gemini work, their key features, and what they mean for their future AI. Don't forget to like, subscribe, and hit the bell icon to never miss any update from Simply Learn. So, what are the large language models? Large language models like CH GPD4 generative pre-trained transformer 4 O and Google Gemini are sophisticated AI system designed to comprehend and generate humanlike text. These models are built using deep learning techniques and are trained on vast data set collected from the internet. They leverage self attention mechanism to analyze relationship between words or tokens allowing them to capture context and produce coherent relevant responses. LLMs have significant application including powering virtual assistant, chatboards, content creation, language translation and supporting research and decision making. Their ability to generate fluent and contextually appropriate text has advanced natural language processing and improved human computer interaction. So now let's see what are large language model used for. Large language models are utilized in scenarios with limited or no domain specific data available for training. These scenarios include both few short and zero short training approaches which rely on the model's strong inductive bias and its capability to derive meaningful representation from a small amount of data or even no data at all. So now let's see how are large language model trained. Large language models typically undergo pre-training on a board all encompassing data set that shares statical similarities with the data set specific to the target task. The objective of pre-training is to enable the model require highlevel feature that can later be applied during the fine-tuning phase for a specific task. So there are some training processes of LLM which involves several steps. The first one is text prep-processing. The textual data is transformed into a numerical representation that the LLM model can effectively process. This conversion may be involve techniques like tokenization, encoding and creating input sequences. The second one is random parameter initialization. The model's parameter are initialized randomly before the training process begins. The third one is input numerical data. The numerical representation of the text data is fed into the model of processing. The model's architecture typically based on transformers allows it to capture the conceptual relationship between the words or tokens in the next. The fourth one is loss function calculation. A loss function calculation measure the discrepancy between the model's prediction and the actual next word or token in a syntax. The LLM model aims to minimize this loss during training. The fifth one is parameter optimization. The model's parameter are registered through optimization technique. This involves calculating gradient and updating the parameters accordingly gradually improving the model's performance. The last one is iterative training. The training process is repeated over multiple iteration or epochs until the model's output achieve a satisfactory level of accuracy on that given task or data set. By following this training process, large language model learn to capture linguistic patterns, understand context and generate coherent responses enabling them to excel at various language related task. The next topic is how do large language models work. So large language models leverage deep neural network to generate output based on patterns learned from the training data. Typically a large language model adopts a transformer architecture which enables the model to identify relationship between words in a sentence irrespective of their position in the sequence. In contrast to RNNs that rely on recurrence to capture token relationship transformer neural network employ self attention as their primary mechanism. Self attention calculates attention scores that determine the importance of each token with respect to the other token in the text sequence facilitating the modeling of intricate relationship within the data. Next let's see application of large language models. Large language models have a wide range of application across various domains. So here are some notable application. The first one is natural language processing NLP. Large language models are used to improve natural language understanding tasks such as sentiment analysis, named entity recognition, text classification and language modeling. The second one is chatbot and virtual assistant. Large language models power conversational agents, chatbots and virtual assistant providing more interactive and humanlike user interaction. The third one is machine translation. Large language models have been used for automatic language translation enabling text translation between different languages with improved accuracy. The fourth one is sentiment analysis. LLMs can analyze and classify the sentiment or emotion expressed in a piece of text which is valuable for market research, brand monitoring and social media analysis. The fifth one is content recommendation. These models can be employed to provide personalized content recommendations enhancing user experience and engagement on platforms such as news website or streaming services. So these application highlight the potential impact of large language models in various domains for improving language understanding automation. In this tutorial we will learn about open gen AI agents. So, generative AI agents are advanced AIdriven systems designed to autonomously perform tasks, generate content, and assist in decision making by leveraging large language models or also known as LLM and machine learning algorithms. Imagine an AI agent automating repetitive processes, enhancing productivity, and providing personalized recommendations across various domains, including customer support, marketing, and software development. By understanding context and generating human-like responses, they improve user engagement and streamline workflows. Businesses benefit from cost savings, faster problem solution, and improved efficiency. Such are a few out of many use cases of AI agents. Today, we will cover the OpenAI gen AI agents and their facts, their challenges, and future implications. That's it. If these are the type of videos you'd like to watch, then hit that like and subscribe buttons and the bell icon to get notified. Open AI gen AI agents. So first let's get started with the agenda for today's session. We will have a brief introduction to generative AI agents. Followed by that we will understand the core components of Gen AI agents. Then how generative AI agents work, architectures and deployment, enhancing AI agents with external tools, challenges and limitations of gen AI agents, future trends and advancements in AI agents will be the last one. So I hope I made myself clear with the agenda. Now let's get started with the first part which is introduction to generative AI agents. So basically what exactly do you mean by the term generative AI agents? So an AI agent is just an LLM that can take actions and it can do so autonomously without any human supervision and now also reason about the tasks. An AI agent is software entity that autonomously interact with environments, make decisions and execute tasks. Generative AI agents extend by leveraging large language models like open GPT to generate intelligent responses and perform tasks beyond rule-based automation. Now we will also look into the evolution of AI agents. Traditional rule-based AI agents had limited to predefined rules like if you define a process and give the steps it will just perform them and complete it. It will not take a step beyond it. And machine learning based AI agents can learn from data but need structured inputs. And then comes the generative AI agents which are capable of understanding, reasoning and generating humanlike responses autonomously like they understand the environment and learn through it and take their steps based on the principle they have been trained for and also they will provide the reason for which they have taken a certain step. Now moving ahead why generative AI agents matter a lot. So there are three key points to consider. Firstly, efficiency. They automate repetitive task with minimal human intervention. Next is high scalability. They handle large volumes of work. Example customer support content generation and many more. And finally, intelligence. They can adapt to different use cases dynamically. Nobody has to write rules for them. Now let's focus on Open AI and its latest release of agents. So, OpenAI just released two new groundbreaking agents, the operator agent and deep research agent. These two releases are absolutely massive for all AI agent developers. So, operator is the first AI agent released by OpenAI and what it does is it mimics human actions on your browser. So, it means it can scroll, type, click and navigate Google web pages for you just like any other human would. But more importantly and what many people have missed out is this agent is actually based on a completely new model called CUA which stands for computer using agent. So this agent unlike standard GPD models wasn't trained to just output text. You like write a prompt and expect an output in the text format. Instead of that it can go beyond it. This agent was actually trained to output mouse and keyboard clicks, which is exactly what makes this agent so powerful in your browser. But not only that, it can also reason about every single action. This is also something that we haven't seen before. This agent doesn't just perform actions in your browser. It actually reasons about every single step and what it should do next. Just like we discussed before about the generic overview of generative AI agents, same as the open AI operator here, it takes the step based on the environment and also it will give you a particular reason for which or based on which it took a certain action and it also asks you what to do next. Let's say if you want to book a ticket, it'll navigate to all the websites. It can go through some uh better flight options for you. Let's see the low cost ones, the low uh hours of flight ones and best timings and then finally selects the best one for you based on your recommendations and asks you if it should proceed with booking or not. So that's how it works. Now of course there are some limitations as well. The first limitation is it has still a lower bit of accuracy. is really high compared to other projects that we have seen in the past where people just prompt GPD models to perform actions in the browser that can browse web using GPT4 vision capabilities but it often struggles with other models. So whenever there was a model on a web page like some other random model available working readily on a web page it would just accidentally pop up and it would not be able to do anything because it would just get confused. it would not be able to close it and then basically you would have to get involved yourself and close this model manually. And the second limitation is high cost. Right now this particular operator agent is only available on $200 plus plan. So we can probably guess it's going to be pretty expensive in the API. Also I do think that if you are focusing on the right thing then costs are not going to be an issue. Now let's talk about the second agent that open AAI deployed. Now the second agent which was deployed by open AAI is called as deep research. So this second agent released by OpenAI was trained to perform comprehensive research. It can search the web. It can pull up the necessary resources and then it can compile all that information into a really comprehensive document or maybe a report as well. The groundbreaking thing about this agent is it's powered by the new 03 model. So again, similar to the operator agent, but it doesn't just search the web and provide you with results. It actually reasons about every single source and then it thinks about what other information it needs to find out next. And it's really incredibly powerful. Deep research doesn't just output the most average thing on the internet just like some other standard GPD models. It would compile novel insights. Now let's move on to the real world applications. There are a wide variety of real world applications. First one is customer support AI agents which helps you in automating help desk responses. Next powerful AI assistance like co-pilots for professionals. Third one is research and analysis agents for data summarization literature reviews and many more. Fourth one is creative writing and content generation automating social media post emails etc. And lastly, AIdriven code assistants like GitHub copilot, OpenAI codeex and many more. Now let's go through the next part which is core components of these generative AI agents. Let's begin with the foundation models. So the foundation models are GPD4, Dolly and Whisper. GPD4 for natural language processing, Dolly for image generation and Wesper for speech to text conversion. The next one would be prompt engineering and context handling. So here designing effective prompts for AI agents to generate relevant responses using few short and zeros learning for adaptability and lastly implementing context aware AI agents that remember past interactions from the user. Now next would be the memory and long-term context storage. AI agents typically process short-term prompts but memory based approaches for example lang chains memory module enable long-term reasoning as well. Now storing and retrieving user specific data for personal responses is also one of the priority. Next will be the tool use and API integrations using external API to fetch real-time data for example weather, stock prices, news, etc. and integrating databases for structured information retrieval. And proceeding further, we have autonomous decision making. AI agents evaluate multiple options before making a decision. Planning and reasoning techniques to enhance responses is a priority for them. Proceeding ahead, we will discuss how generative AI agents work. Starting with the first step, input processing and context awareness. AI agents analyze user input, detect intent and extract entities. For example, in a travel booking AI agent, it identifies destination date and budget from user queries. Just like the example we discussed before, right? You provide the point A and point B, the best time for you to fly and the best number of hours that you focus on flying and lastly the destination, right? All these things will be evaluated in real time, the cost, the number of hours, the flight, the options provided, the amenities provided on the flight and all those things and also budget as well and based on that they take a decision. Now proceeding ahead generating intelligent responses. So here R A or also known as retrieval augmented generation. AI searches external databases responding. Next we have chain of thoughts reasoning or also known as coot reasoning and AI breaks down problems into fine steps. And last step is fine-tuning the models. Customtrained AI for specific industries. For example, if you're preparing an AI agent for a dedicated sector like healthcare, legal, customer support, based on that, you will fine-tune your AI agent or model. Now, next is decision trees and multi-step reasoning. AI can simulate a thought process, breaking down complex questions into manageable steps. For example, whenever you open Deep Seek or whenever you open Open AI Chart GPD for the most premium version, you will write a prompt and there will be an option called think, right? If you just click on that, it will explain you the thought process it is going through and how is it planning on giving a reason for you, right? And followed by that we have handling uncertaintity and error correction. AI self-corrects through confidence scoring and human feedback loops. Last step in this particular stage is fine-tuning and reinforcement learning or also known as RLHF. Human in the loop training to improve responses and AI learns from real world interactions and adapts over time. So it's basically called overtime learning where it adapts to the real world environment and learns things and provides the best feasible solution. Now the next stage is architecture and deployment of AI agents. So in the first one we have standalone versus multi- aent systems. So firstly the standalone AI agents they perform task independently. For example chat GPD multi- aent systems in the other hand work together to achieve a complex task. For example one agent for summarization another for execution. Now next in this particular stage is agent orchestrator modeling. A central AI agent coordinates multiple AI sub aents. For example, in an AI powered research assistant, one agent gathers all the data that is required. The second agent summarizes all the information collected and the third agent formats the output. Followed by that we have cloud-based AI agents and edgebased AI agents. Cloud-based AI agents compute, for example, OpenAI API, Azure AI, and edge-based AI agents run on local devices for faster processing, for example, Tesla's AI in self-driving cars. Followed by that, we have using APIs to build AI agents. OpenAI's API for GPD4 powered by AI agents, Langchain framework for AI agent orchestration, and Llama index for document-based AI agents. The next stage we will discuss about enhancing AI agents with external tools. In that the first step is integrating with databases. AI agents can query SQL and non-SQL query databases for structured data retrieval. Proceeding ahead we have connecting with web scrapers and APIs. AI agents can fetch realtime data using tools like beautiful soup scrappy and open AI plugins. Next ahead, using vector databases or context recall. F AISS Chroma DB store long-term memory for AI agents. For example, legal AI agents retrieve case law from vector embeddings. Lastly, we have autonomous agents and workflows. AutoGPD and baby AGI are best examples. these AI agents that self-improve by learning from previous task and true AI which is a set of multiple AI agents collaborating to solve a complex problem. Now the main part challenges and limitations of generative AI agents in that the first one will be the ethical considerations. AI bias and fairness concerns are the primary ones addressing AI hallucinations which is false information generation and then we have computational cost and efficiency which is a major concern amongst all the industries. Large AI models require significant computational power. Strategies to optimize AI agent efficiency will also considered to be expensive to hire many training models or also human programmers. Followed by that we have security risks and data privacy. Risk of data leaks or data breaches and AI jailbreaking is one of the highest security threat. Implementing secure AI architecture is one of the primary solutions that you need to focus on. And lastly, we have legal and compliance issues. This would major be concerned from the enterprise level. GDPR, HIPP, AAA and AI regulations compliance is mandatory and AI liability in decision making is also important. And now the last stage of today's discussion which is future trends and advancements in AI agents. Starting with the first one, AI agents with emotional intelligence or also known as effective AI. AI is capable of detecting emotions and adapting responses accordingly. For example, if you have designed a therapeutic AI chatbot for mental health. Now the next one, self-improving AI systems. So metalarning and AutoML are a few examples. AI agents that learn from new data autonomously. For example, AI adjusting marketing campaigns based on customer responses. And now the next one, AGI and road map to super intelligence. Well, this always ends up with question. Will AI agents evolve into AGI or also known as artificial general intelligence? And are the implications of self-arning AI real? This is the major question. And lastly, the role of AI agents in the future of work. Now we can see there are a lot of AI agents like copilot which will assist humans in their day-to-day task whether it is business healthcare or education and also you have a lot of automation and transformation happening in the IT industry which will reduce the work hours and stress on the employees by automating all those mundane tasks by the help of co-pilots also the generative AI agents. Hey everyone welcome to the simply learns YouTube channel. In this tutorial, we will learn about the MCP. Anthropex MCP or also known as the multi- aent collaboration protocol is a groundbreaking framework designed to enable seamless cooperation and collaboration between multiple AI agents. Unlike traditional single agent ecosystems, MCP allows agents to specialize, communicate, and coordinate effectively to solve complex tasks. This leads to a faster decision making, enhanced problem solving and more scalable AI systems. One of its key strengths is structured communication where agents can negotiate rules, share knowledge, and adapt dynamically. With MCP, developers can now build AI ecosystems that behave like more intelligent teams than isolated tools. It marks a significant leap towards building truly autonomous collaborative AI applications. Now in today's session, we will actually try to work on one and we will be building a web scraping AI agent using MCP protocols. That said, if these are the type of videos you'd like to watch, then hit that like and subscribe buttons and the bell icon to get notified. Now we will begin with the agenda for MCP tutorial for beginners. So firstly we will have a quick briefing of what exactly is an MCP or as it is called multi- aent collaboration protocol. Followed by that we will understand some basic protocols and standards in communication. Next we will try to relate MCP to the understanding of protocols and standards. Followed by that we will get started with our demonstration on developing a web scraper agent using MCP. And for that we will be needing NodeJS cursor AI and firecroll API. So I will be walking you through those demonstration steps one after the other. And this will be a very beginnerfriendly tutorial and practical demonstration on MCP. Now that I've made myself clear with the agenda, let's get started with MCP. So what exactly is MCP? As MCP on their web page says, MCP is a model context protocol or multi- aent collaboration protocol is an open source protocol that standardizes how applications supply context to large language models. Similar to how USBC unifies hardware connections, MCP offers a consistent way to integrate AI models with various data sources and tools. Now let's go to the official documentation of MCP to have a broader overview. So as they say MCP is one such tool if you remember N8N and all the AI tools and data sources that you used. For example, if you wanted an AI travel agent as we have an already existing tutorial on our channel, you can go through it for a brief overview. So you might be needing an email writing agent and you might be needing an airport code searching AI agent, hotel searching AI agent, flight booking AI agent and a lot of agents connected to one another through a complex network using an ATN platform. Right now in this particular MCP that will be minimized. So you will have a standardized approach where you just eliminate all those boilerplate coding agents, manual searching for databases, manual search for knowledge bases, manual search for agents which are specifically designed for the purpose. You just explain your initiative, attach the API keys and MCP takes over and gives you a fully furnished and finished product within no time. That's the main approach of MCP. So here you can see an overview image where we have the parts MCP host clients servers local data sources remote services and local data sources. If you wanted to go ahead with local data sources or local knowledge base you can go ahead with that. But in case if you wanted some open source then you can also go ahead with that. So here we have all the quick starts and guides and example and tutorials to how to begin with that one and all the explore pages that you want to go ahead with if you wanted to have a deep dive into core architecture the resources available promps tools transport samplings and a lot more. And in case if you wanted to contribute your dedicated resources for the community you can also go ahead with that and you can also have some support and feedback for MCP. Now that we have a brief understanding of MCP and still clueless about what exactly is MCP then let's go ahead with some standards and protocols so that you can get a template and you can relate MCP to that template and understand what's the role or importance of MCP in AI agent development. Now understanding the standards and protocols. Now let us imagine that you get a team of people to work on a project and you are the manager that leads and directs this particular team for a purpose. Right? Now the complex problem no one is friend of anyone. Everyone is new to everyone on the team and each and every person in this team belongs to a different country and does not have a common language in between. For example, person A speaks Mandarin, person B speaks Hindi, person C speaks French, so on. There is no common language between any one of those. Now, imagine each and every one of these team members as an AI agent or a tool working for the purpose. In this scenario, you might have to hardcode the ports of communication instructions on how to do and you need to go on to each and every single AI agent and explain the stuff in a very dedicated and a specific way. This could take a very long time. In the same way, if you relate this to a person, then you might have to reach to every single person and speak to them in their own tongue and explain them the procedures and the steps to be taken to finish the task. Now, how would you do that? It would be a lot complex. Mandarin, you might not even know, French or a different language. And in that particular scenario, you might be stuck. What if there was one single language that everyone knew, for example, English? Now, you can get all of them together and explain the stuff in one single call and close it. And what might be so wonderful is that you'll not even waste a lot of time and there is no hard coding. There is no manual effort there. Right? In the same way, if you relate this to a web server, then you are the client. You have the server and you just let's say you're working on Facebook or Instagram. You try to post something, you try to write a command, you try to like something, right? That is the activity. And these particular activities will go through four different types of requests. One is the get request, put request, post request and last is the delete request. Right now if you are connected to a single server this might not be as tough as possible but if you are getting connected to multiple servers then you might be requiring different protocols every single time now this can be a combination of two parts HTTP and the URL that you're trying to get connected to right now to eliminate the efforts between multiple servers and multiple requests at one go you have one single tool called as API And the most popular API is the rest API which has a standard way of communicating. So here the rest API works as the language English that we discussed before. In the same way, it has maintained a standard where each and every web client can have his communication to his dedicated server or multiple servers in one go using one single language or medium or standard that is rest API. Without REST API and the standards and protocols, it might be very confusing and you might end up in a complex situation which might not yield in the type of result you're expecting. Now if you can relate the development of multiple AI agents for a single purpose to the web communication and a meeting you also expect if there were a template for development of AI agents which could make the process less complex just like rest API did in a web service and just like English did in a mid team meeting then MCP comes into the picture. MCP is your template for it right? So just like your rest API and the meetings you have MCP. Earlier without MCP we used to have one tool called NA10 where you had all the tools connected to it and all the LLMs programmed to it and you had to do this hardcode coding for each and every tool that you have included in the network and each and every agent that you have related to it. You had to specifically instruct, write down steps, develop a persona to your AI agent, right? There was a lot of stuff. But if you eliminate this particular complexity with one tool, then that would be your MCP. You just had to include MCP and tell your purpose and your job would be done. Now that you have a brief understanding or a general overview of what exactly is MCP, now let's go ahead and build one. Now the steps would include the following. For the project that we are building which is a web crawler or web scraper we might need the help of NodeJS. So firstly we will install and set the environment for NodeJS. Followed by that we will be needing an editor. So for that we will be choosing the cursor AI editor. And next we will use an API. So for that we will be using the firecrawl API. And lastly, we can get started with building one once we have all the required resources. So let's get started. So as we discussed before, we will get started by installing NodeJS into our operating system. So we're working on Windows operating system. So we might have to go ahead with Windows. So I'll choose Windows and the architecture will be the one which I'm working with. So that would be x64. But in case if you are working with a different operating system say Mac OS then you might have to select ARM or any suitable architecture that you can be compatible with for Linux you'll have to choose with the same and so on. So since we're working with Windows so Windows it is and X64 architecture and click on that Windows installer MSI file. So it will be getting started to install and you'll have to click on that Windows installer MSI file and it'll get downloaded. While it is getting downloaded, let's go to the next web page and download and install the cursor AI that we will be needing for this demonstration today. You'll have to login, create an account, and login. And automatically, it'll give you the version that suits our operating system. Since we're on Windows, it gave me a Windows operating system. Just sign up or login. So, since you've already an account, just log into your account and confirm it's that you and followed by that, you'll have uh installed your cursor AI. Choose your theme. theme. I'll go with the cursor dark theme. And here you make sure that you have cursor tab, control key, agent, and you can also choose the data sharing option. I'll go with the privacy mode in case if you want the other one. Go ahead with that. And uh here you have a few review settings. And there you go. Now you're on the new window for cursor. And meanwhile, let's check on the NodeJS installation. I think uh the wizard has already started installing the NodeJS setup. So uh this is the window. Let's quickly uh go through each and every instance that we have gone through so far in this particular window from the beginning. So it started off with some u packages that we would be needing. So uh yeah chocolatey and uh KB299. Yeah. Further we have and the last one is Python 13 right version 3.13.3 is approved so far so good I think we can now get started with the next step so here we are on the cursor AI so we might have to uh get started going through the cursor AI settings to create our new MCP so there are quite a few methods that you can try out to reach to settings one is you can go to the file and there you may have to choose the preference senses and then you have the cursor settings. You just can use that vertical list or you can go ahead with the search options as well. So you can use that step or you can also use the other step where you can just go to the search window and search for settings. Here you can see uh MCP on the left window. So let me also walk you through a few other steps. So go to the search window and uh type down cursor AI or cursor settings. I think there has been a discrepancy but not to worry we'll go with the regular approach where we go to uh the file preferences and cursor settings and there you have uh the options on the left hand side go to MCP and so far you don't have any MCP servers because this is the first time you logged into cursor AI so you just have to add new global MCP server which is available on the blue color button just click on that and now you'll have a mcp.json JSON file which is completely empty. You'll have to write down a few hardcoded codes here which are not too complex related to the earlier ones that you have worked on in A10. So you might need a few more u tools here. So go to the fire crawl website. Before that, let's go through some uh awesome MCP servers GitHub link. So here we have almost all information related to the MCP. So let's open a new window and open all these links. What is MCP? Clients, tutorials, community, legend, server implementations, framework, tips and tricks, right? For further information, you can go through all these web pages individually. So back to the homepage here, you have everything you needed. The basic information of what exactly is MCP and further you have some server implementations. Like if you scroll down, so you have awesome MCP services. This is where the beginning is and next you have server implementations, aggregators, art and culture, coding execution, right? Customer data platforms. So here you have aggregators, art and culture dedicated web spaces. So and you have also cloud platforms that you can work along with. And moving further you have uh code executions, coding agents, command lines as well and communication related uh stuff. So you can go along and perform your operations, customer data platforms, data platforms and much more. Let's uh close all these windows and get started with the next stage where we need an API. Right? So for that uh the project that we're working on which is uh web crawler, we will be needing an API from one of the uh providers which is fire crawl API. Okay, that's the one firecrol mcp. So that's the one. Click on that particular link here to generate an API. First you might have to sign up with fire crawl. So uh before we sign up let's go through the website of fire crawl from the beginning. So that's the template code that you'll be working on. And also here we have a wide variety of uh uh crawl options, media parsing options, smart weight actions, right? Reliability first and all those things. And also we have the pricing details. For now we'll go ahead with the free plan. And uh yeah, let me add down my details. My email address and you might also have to add a password. So let me type down something. Yeah, so the password is not too strong enough. That's a good thing. They have a really strong website. So uh that's a sign you need to have strong passwords and also make sure that you don't share your passwords and the API keys that you generate using firecrol. Now that my password is a little too longer, I think it should be excess. Yeah, it's accessible and is a successful account generation. Now you need to check your email and confirm that uh login credentials are right and everything. So once you are signed up, let's go through another website for generating our API keys. So fire API scroll a little bit. So API documentation. So here you'll find some documentation related to fireroll and all the details that you need to go through for API key generation. So fireroll mcp server. Yeah, that's the website that you're looking for. Firecrol MCP server. That's the one. And here you have overview content tools and other resources that you're looking for. Now this is the template server configuration that you'll be needing. So make sure that you have an eye on it. And now you will also have uh the details that you might have to go through like basically it'll have uh the steps that you need to follow in the content page. Click on content. There you go. So it has all the uh features and uh details that you need to go ahead with and tools. You have the crawling extraction scraping tools that you can work along with fire crawl and uh few comments here. Now yeah the email has not been confirmed yet. Let's try to resend and uh reconfirm the email so that we just sign up or login and create our API keys as soon as possible. So there you go. I have verified my email in my phone. So I'll just enter the credentials and sign in. Successful. You are signed in. On the left hand side you can see the overview playground extract u activity logs usage API keys and all the settings. So here you can see my API key has been by default hidden. So I'll not visualize it now. So you have two options either view or copy. So we will be copying it for the first time to be on the safe side. So uh since we are new to this particular website, so we don't have any activities registered on this particular web page. So now we have the API keys. Now let's go to the content page to verify the steps that we need to follow. So we need to install npm and we need to go to the cursor AI settings and add a new MCP server which we already did. So we just need to copy the template code and we just need to place it there. And also don't worry about uh where exactly to place your API keys in the template. We have that place. So we are back on the cursor AI. Now we will replace this template code with the code that we just copied. And remember, if you're using Windows, try cmd/ C. Okay, if you're not on Windows, you can eliminate that uh cmd/ C. Uh but if you are working in Windows, then you definitely need that cmd/ C. Okay, let's erase this and replace that. So I'll name my tool as fire crawl and I'll go to the new line where I'll write down my command and the command will be the code that I just copied from the fire crawl window. So let me type down the command in double quotes in new line and I'll paste the command here which is the code that we copied from the back window. Ctrl V. There you go. So the white stuff which is in cmd/ C. So that's the one you need to focus on. If you are working on Windows that you definitely need it. If you're not working on Windows then you can just omit it. Now uh there might be some concerns over the codes. Don't worry. Now to crossverify you can just go back to the firecrol homepage and uh crossverify it with the template for configuration that they have uh made us available. Let's go back to the firewall window and at the same time you can see your API key here. So that's where you need to place your API key. So here we are on the firewall window. You need to copy the API key and place it in the place where they have mentioned your API key. So if you have cross verified with the configuration code that's fine and now you can just uh check the double quotes. So that's the one you need to focus on. Let me uh kind of uh if there is a way to zoom it in. Let me try to zoom in. Yeah, that's the one. cmd/ C. If you're on Windows, you definitely need this. If you're not on Windows, you can just choose to permit it. Now, let me just do a one last final check. I think I missed a double quote somewhere. Yeah, I need to remove it from this place and add it to the end. Just a second. Okay, remove the double quote at that place. Now, it should be after MCP. And once everything is done, you can just quickly uh choose to save and close this particular window. Check the distance between the API key and the double amp places. That's fine. And saving is really simple. Choose those three dots. Close. Save. And once that is clicked, it's all done. You get to, you know, choose to close this particular mcp.json file and get back to your agent. So once you have closed it, you will have your new server that you've just created in the cursor settings. But in a few times you get uh no tool connected, no server connected or there might be some errors. So in that case you just need to cross verify the npm installation, NodeJS installation and uh check all the files that are there or not. But even if everything is installed and you can see uh you know the versions available and if you're still seeing this or watching that particular error then don't worry just close the cursor AI and restarted or you can also choose the refresh option which is right available there. So I've closed and restarted it and the same procedure go to file preferences cursor settings and there you'll see the MCP models and there you'll find it right there. Correct. Now close it and go to the window where you'll find the uh server. Yeah, that's the one. And here you choose the agent. Make sure that you choose the agent instead of anything else. Not manually, not ask. I choose the agent. And now I'll write down a simple prompt. Search for simply learn on Google. Now it should give me results which it finds on Google. Maybe it's a homepage. Maybe it's a social media handle like LinkedIn or YouTube or Instagram. anything that it finds related to simply done only the term simply done. So here you can see it has given me three different results. First one is simply website second one is simply done YouTube channel and the third one is LinkedIn page. So let's get started with the official web page. So click on the simply.com. Now uh yeah it's asking permission open. Now we are on simply learn homepage. Similarly let's go back and select for the YouTube uh page. So that's the one. Open again. Get permissions. So you're on a simply on YouTube page. Now I think it might ask for a login for LinkedIn. But anyways, let's try to check if it can open without logging in to LinkedIn or not. So yeah, you can see I've logged into the YouTube. Yeah, it is asking me to login. So let's quickly log in. It might take a couple of seconds. So there you go. I've logged into LinkedIn. Now open. So I can see simply learns homepage getting loaded in a second. Yeah, there you go. On LinkedIn we have opened the simply learns homepage successfully. Introduction to agentic workflow from simply learn. So before moving on to creating the agentic workflow let's first understand what is AWS bedrock. So AWS Bedrock is a fully managed AI service that helps developers build and use AI powered applications without needing to worry about setting up or managing complex infrastructure. So it gives access to pre-trained AI models from top providers like Amazon, Anthropic Plot, Meta, Misti, etc. So with Bedrock, you can create chatbots, text generators, summarization tools, and other AIdriven features without needing to train models from scratch. It also lets you customize AI models using your own data, add guardrails to ensure your safe and appropriate responses and connect knowledge bases so the AI can provide more accurate answers. So since VRO is part of AWS, it also works smoothly with other AWS services making it easy to integrate AI into your existing applications. So whether you're working to automate customer support, generate content, or enhance decision making, VR provides a simple and scalable way to bring AI into your business. So now let's talk about serverless agentic workflows. So as AIdriven automation continues to evolve, business require intelligent agents that can understand user queries, retrieve relevant information and take meaningful actions. So Amazon Bedrock introduces serverless agentic workflows enabling organizations to build AI powered assistance that interact with users, fetch realtime data and execute business tasks all without manual infrastructure. So in a traditional AI chatbot, the system typically responds to queries based on predefined answers or static knowledge. However, with serverless agentic workflows in Amazon Bedrock, AI agents can dynamically retrieve information and trigger actions, making them significantly more powerful than basic chatbots. So now let's explore how these workflows function, their key components and their real world applications. So when a user interacts with an Amazon Bedrock agent, the following steps takes place. First one is user input processing. So the agent receives a query from the user such as I need to return a laptop I purchased last week. Can you help me? And the foundation model processes the input, extracts key information and decides the next step. And the next step is information retrieval from a knowledge base. So if additional details are needed, the agent queries a knowledge base to pull relevant information. And the third step is action execution via AWS Lambda. So if the agent needs to perform an action, it calls an AWS Lambda function to interact with backend systems. So the function retrieves the order details uh verifies eligibility and processes the return request for example. And number four is generating and sending a response. So the agent gathers the information and constructs a natural language response for the user. And finally the session continuity and follow-up. So if the user asks them the question, the agent continues the conversation while maintaining context. So the entire workflow runs on AWS infrastructure allowing businesses to deploy AI powered assistance without worrying about managing servers, scaling resources or or handling background execution. So here are some of the key components of serverless agentic workflows. So number one is Amazon bedrock agent. So the core AI assistant that interacts with users, understands their queries and determines the best course of action. So it basically uses pre-trained foundation models from providers like Amazon, Mistral, etc. to process and generate response. And then it has foundation models. So these are large language models that power the agent's natural language processing capabilities. So the model helps the agent understand questions, retrieve information and generate responses. And then there are action groups. So action group define what the agent can do beyond answering questions. They allow the agent to execute functions, fetch data, and interact with APIs. So for example, retrieving customer orders from a database, processing refunds using an e-commerce system or creating support tickets in a help desk platform. So each action group is mapped to an AWS Lambda function which handles the actual task execution. And then we have AWS Lambda for business logic execution. So since bedrock agents are serverless, AWS Lambda is used to execute backend logic without managing infrastructure. So for example, a lambda function can look up order history, verify eligibility and initiate a return process. And another lambda function can check support ticket statuses and provide updates. And fifth, we have knowledge bases for data retrieval. So a knowledge base helps an agent retrieve factual information from structured databases, FAQs or documents. And then finally, we have guardrails for AI safety and compliance. So guardrails ensure the agents responses remain safe, ethical and appropriate by filtering out harmful, sensitive or restricted content. So here are some of the real world use cases of agentic workflows. So number one is for customer support automation. So a company can deploy a bedrock powered AI agent to handle customer service tasks like answering FAQs using a knowledge base, retrieving order details via an action group plus lambda function and initiating return requests and processing refunds automatically. It can also be used in IT help desk support. So an internal AI assistant can help employees troubleshoot technical issues such as checking server uptime using AWS Lambda or resetting passwords via an action group request or submitting IT tickets automatically in support system and it can also be used in HR virtual assistant like for example a bedrock agent can be used to assist employers with HR related queries like fetching leave balances, processing leave requests or providing company policy details. So basically serverless agent workflows in Amazon bedrock unlock a powerful scalable and cost effective way to build AIdriven automation. So these workflows represent a significant step forward in AIdriven automation like allowing organizations to deploy intelligent actionoriented AI assistance that go beyond simple conversation. So now in the coming chapters we will see how to create agent with Amazon Bedrock, how to connect with CRM and perform calculations and how to read an FAQ manual and a short Amazon Bedrock console walk through. So let's get started. Okay, but before we start in this course we have used the claw 3 haiku model. This is one of the latest AI models and here is a quick look at its pricing. For claw 3 hiku the cost is $.25 per 1 million input tokens and 1.25 $225 per 1 million output tokens. So throughout this course, we'll be using cliha as a primary AI model. So all right, let's proceed with the course. Hello all, welcome to our first chapter. So in this chapter, we are diving into Amazon vetro to build a simple cloud-based agent. You'll also learn how to invoke the agent and trace its workflow through traces, giving you a clear view of how it happens. So let's get started. The first step just like in every chapter is importing the bottle 3 library. So this is basically AWS official SDK for Python which you'll be using throughout the chapters. But before we start we have to you know set up an environment to run the code. So first let's set up the environment report. I'll be explaining you everything shortly. Is equal to OS dot environ. So basically you know this code will reset the environment and prepare the resources. So let's just run this. Okay. Sorry there is a syntax error. Let's fix that. Yes, as you can see it is resetting the environment. So the agent reset process is completed. Lambda reset process is completed. Guardrail reset process completed. We have discussed all these in the you know introduction chapter. So the environment is set. So now let's start working on it. As I already mentioned before, let's import the bottle 3. input 23. Now the next step is we have to create a client object. So basically we create a client object to interact with the different AWS services. So let's do that. So bedrock agent is equal to three dot client service name is drop agent and we'll also you know provide the region name. So region name is west 2. Now next we'll use this client to connect to the AWS and set up our agent there. So for your information this code is configuring in a cloud-based agent that operates at production skills. It won't be running in our notebook right now. So let's get started. So let's start by creating the agent. So this is actually pretty straightforward. We'll be using the bedrock agent client and calling the create agent function. So let's write the code. So first we'll be using the bedrock agent client and calling the create agent function. So now let's you know pass the parameters. Let's set up the parameters. So the first thing is we need to give our agent a name. You know we should call it something. So one thing to keep in mind is this name isn't something that the agent itself will recognize. It's just for our reference. So for example, I will name this one as laptop customer support agent. So now let's move on to the other parameters. First things first, we need to give our agent a foundation model. So basically this means selecting a specific large language model from Amazon bedrock which will handle the language understanding for how agent operates. So we'll paste the model ID over here. So the foundation model is anthropic cloud 3 haiku. So now there are multiple models available and you can find the full list in the as documentation which will I'll be showing you know further chapters. So here we have set up with anthropics cloud 3 hiku. This is very small, fast and highly capable model. So the next step is we need to define the agent's behavior by providing it with instructions. Now these are also known as system prompts. So let's give the instruction. So we have given it the instruction that you are an advanced AI agent acting as a frontline customer support agent. So in this case the agent is acting as an advanced AI assistant for the frontline customer support especially for laptop businesses. Now that we have set up the model and instructions there's just one more thing left for the configuration that is the resource code. So predefined IM rule has already been created for us in this environment and we are pulling in its AR and AR means Amazon resource name from an environment available. So basically this rule defines the security permissions for our agent. basically dictates which services the agent can access. So in this case we are keeping the things secure and minimal following the principles of least privilege meaning the agent only has access to necessary language model and nothing else. So let's set up that and yes we are good to go. As you can see we have loaded this into a variable create agent response. Let's just check it. Yes. So as you can see our agent is ready. It is created. So now you might have noticed that we haven't assigned any tools yet. So in Amazon Redrock these are called actions and they belong to action groups. So for now we have built a basic AI powered laptop support agent and the next step is learning how to invoke and interact with it. So let's see how to do that. So as you can see our agent status right now is in creating state. We have the agent ID. We have the agent name. So basically what happens next is the agent moves to a state called not prepared. So basically this is the full process the agent goes through. First state is creating where the agent is being set up which is happening right now. Then the second state is not prepared where it exists but isn't ready yet. And the third one is prepared where it's ready to use. So once it's prepared, we can create an alias for it. Alias is like a stable version of the agent that we can use in production. So because it lets us control updates instead of changing the agent directly, we can manage different versions and keep things stable. But before doing anything else, we need to wait until the agent reaches the not prepared state and that's when we can move forward. So now let's go ahead and store the agent ID in a variable. So basically I'm going to pull it from the response we created earlier from the agent field where the agent ID is located. So let's do that. So agent ID. This is the variable create agent response agent agent ID. So this is really going to be very useful for us. Now we have to set up a few helper functions here. So for example, they'll help us wait for certain processes like transitioning to the prepared state. So we don't have to handle everything manually. And since we are working on laptop customer support, these functions will make it easier to manage tasks like checking the laptop's warranty status, troubleshooting hardware issues, and guiding customers through software fixes. So first let's import the helper function. So from helper let's first write the helper function and then I will explain the code to wait for agent status agent ID agent ID A target status is not prepared. So as you can see that the agent has reached the not prepared state which is actually a good thing. That's exactly what we wanted. So it has basically transitioned from creating state to not prepared state. So now the next step is to prepare the code. So it's ready to go. So let's do that. I want it to be prepared state. So as you can see now the agent is preparing. Agent state it is preparing. So we are using the bedrock agent which is a client. If you remember we set up earlier and we have called the prepare agent function. So basically we need to tell it which agent we are working with. So we'll pass in the agent ID also. We have passed the agent ID which we had already stored in a variable and thus it has reached the preparing state. Now our target is for prepared state. So let's just take this code and our target status is prepared. Let's see. So as you can see the agent has reached the prepared state. But if you have noticed as I showed you it did briefly show preparing before it switched to prepared state. Now as I mentioned earlier the next step is to create an alias and that is because we want to invoke the agent and have a conversation with it. So since the agent that we created is now prepared we can go ahead and set up the alias. So let's do that. So let's move on to creating an agent alias using our bedrock agent. Let's do that. Create agent is Sorry. Response agent create agent alias. Yeah. So to do this we need to provide a few key details. So first we have to specify the agent ID so that the system knows which agent we're working with. Along with that we need to assign a name to the alias. So this name can be anything. For simplicity we'll just call it my agent alias. So let's do that. First let's specify the agent ID. Then let's set the agent alias name which is my agent in this. So now once we have these details we'll store the response from this operation in a variable right. So this basically allows us to reference it later when we need it. So the process involves calling the create agent alias function passing it both the agent ID and the alias name. So let's do that for that agent sorry alias ID equal to create agent area response. We have given the you know we have passed on the agent alias and agent alias id. So now before executing this there's an additional step to ensure that everything runs smoothly. So we need to extract the agent alias ID from the response and store it in a variable. So this is important because we'll use it to track the alias status. So to make sure that the alias is properly set up we'll use a helper function that waits for the alias status to transition to prepare. So let's write the helper function also. Let's include the helper function also. So yes, this is the helper function. So we have already talked about helper functions earlier. So now let's execute the code and see what happens. So as you can see everything is working as expected. The alias transitions into the prepared state which is exactly what we waited for. Now that the alias is ready, we can actually start using the agent. So with prepare als the agent is now in a functional stage meaning that we can interact with it. So let's move on and see how to put it into use. So now to interact with the agent we need to create a new client specifically for this purpose. So we'll be using b3 to initialize it. So let's create a new client as we have done earlier. As you can see we have you know specified the service name and we have specified the region then we have stored this client you know variable. So let's run this. Now that we are ready to move forward and actually invoke the agent. So for that we have to write a few lines of code. So first you have to import U ID. So we'll be calling the invoke agent method for our bedrock agent runtime client which is essentially all that's needed to start a conversation with the agent. So let's write the code first and then I'll explain it to you. So as you can see this is the message that I've given as a customer that hello I bought a laptop from your store yesterday. there is an issue with the display and I want to return it. So as you can see we have specified which agent it is by giving the agent ID. Similarly we also passed the agent alias ID which we had saved earlier in the process. And finally we gave the message also. So at this stage the agent uh doesn't have any actions configured meaning it would actually be able to troubleshoot the issue. However, it should still generate a response based on whatever logic is available. So, in addition to the message that we have given, we also have given a session ID. We have to establish a session ID to maintain the conversation. Right? So, you might have experienced chatting with the you know customer support agent and you might have noticed that you know there's a session ID and there's a session for each and every conversation. So, once you end the conversation when you're starting a new conversation you know it's separately a new session. So similarly here also we need to establish a session ID to maintain the conversation. See we have generated a unique ID using Python's UU ID module and that's why we have imported UU ID first. So basically this ensures that each session is uniquely identifiable and the conversation history will be stored in the cloud. So we don't have to manage it ourselves. So here we have kept end session false which means this keeps the conversation active instead of losing it after one exchange and enable trace is true which means this and enable trace for now let's keep it as false. Yeah which means this disables request tracing for now but we can explore it later. So let's just end this change response. So as you can see we have run the code over here and as we print the response you'll see that it doesn't display like a normal text response and that's because it comes back as an event stream instead of a simple message. So since we can't just print it directly, we need to extract the actual response from the stream. So to do that, we'll pull the relevant data from the event stream which will allow us to see what the agent actually said. So let's do that. Let's write a piece of code. Event stream is equal to invoke agent response completion. I event. So as you can see when we do this we'll see that there's only one main event returned and this event is called shrunk and inside it we have some data specifically bytes and a response from the agent. Basically, the agent has responded that I'm sorry, but I'm unable to assist you with the return process for your laptop at this time. Please contact our customer support team directly and they will be able to help you with it. So, this message makes sense because right now the agent doesn't have any actions set up, it knows it's supposed to have some functions, but since we haven't defined any, it simply acknowledges its limitations. But how does the agent figure this out? And how does it do the necessary action? So for that to get a clear picture we have to enable the trace mode. So earlier we have enabled disabled the trace mode by keeping it false. So we have to enable it by making true. So this will start a new session. It will reset everything and allow us to have another conversation with the agent and this time with more insights into how it generates it response. So let's go ahead and do that. So as you can see here I have enabled the trace and I printed it here. As you can see we can find some familiar details like they have given the agent ID, the session ID. They have given so many details and I'm just finding the final response over here. As you can see it has given a final response to initiate the return process for your laptop with the displays. Please gather the following information. Date of purchase you know order number of receipt description of issue. So as you can see first it said that you know I I'm I am unable to help and please contact you know uh some customer support agent like you know some human customer support but as you can see now it actually you know gave us a answer that I want you know some details like date of purchase order number or receipt etc so that we can initiate uh you know the return process. So with this in place we are now ready for the next step which is adding actions to the agent so that it can actually perform some useful tasks. So we'll see that in the next chapter. So now that we have a basic agent set up and running, it's time to take things a step further. So in this chapter, we'll connect the agent to external services allowing it to interact with the outside world. We'll simulate integrating it with a CRM system where it can retrieve customer details, log support tickets. So this will make our customer support agent more functional and bring it closer to real world use case. So let's see how it works. So first things first, let's set up the environment. Do that. So as you can see the environment is being set up. This will take some time. Now next we need to import the required libraries. We'll start with mod 23 to interact with AWS services. Import 3. We'll also bring in UU ID which we discussed in the earlier chapter which helps us create unique session ids and also we'll import some helper functions. So let's do that. Import from helper. import. So since our environment variables are already set up, we all set to go. So the agent from the last chapter from the last environment is still running. So we can continue using it. And this time we'll give more detailed instructions because we have added a few extra components. So now let's start testing and see how it works. So firstly I'll generate a new session ID and set up another message similar to what we used in the previous chapter. Uh let's set up the session ID first. Session ID and let's new message. Okay. Let's give my laptop as display issue and I want a refund. So once that's defined, we'll use a helper function to check the responses and see what happens. So we'll call invoke agent and use a print help function to send a message and check responses. So let's do that. Invoke agent. Sorry, invoke agent and print. Again we have to specify our id text message agent als ID is equal to agent als ID. [Music] session. So, as you can see, this is what our response is. You know, this actually looks like a chat format. We said that my name is Leah. My laptop has displays and I want a refund. And the agent has responded by saying that it's sorry, but it doesn't have enough information to process our request for a refund at this time. And you know all the details. So now what we need to do is we need to connect an agent to some other tools to make it more useful. And in Amazon wedrock agents, these tools are called actions and action groups. These allow the agent to perform tasks beyond just responding with text. So the logic behind these actions is handled through AWS Lambda functions. So basically the lambda function serves as the invocation point where the agent sends requests when an action needs to be performed. So when the agent calls this function passes an event that contains key details like agents name, action group name, specific function and in parameters. So let's do that now. Let's write the code first and then we'll explain. So to connect an agent with tools, we need to define an action group using the bedrock agent client. So this action group will allow the agent to perform tasks by interacting with an AWS Lambda function. So first let's create the client like before. So we have done that the service name is bedrock agent region region name. We're not specifying the region right now. Now next step is to set up the action group. So yeah, now let's see what the code is. So we have started by naming the action group. The action group name is customer support actions and this is just for our reference. Now the next thing is we link it to our agent by providing the agent ID from our environment variables and we specify the lambda function that will handle the logic using the ARM. Now after that we have defined the actions like for example inside the action group we define functions the agent can call. For example, the first one is customer ID lookup. So this function retrieves a customer ID based on details like email, name or phone number and the agent must send at least one of these details to proceed. Now the parameters are described in natural language so that the agent understands their purpose. Now the next function, the second function is send to support. So basically this function escalates the issue to a human support team. Like for example, if the you know agent that we created is not able to respond. So it will escalate the issue to a human support team. It requires both a customer ID and a support summary. And if the agent doesn't have a customer ID, it will first fetch it using the customer ID lookup function. So this is what these two uh functions does. Now before sending the request, we have specified the agent version as draft. It means that it applies to the current agent setter. So once everything is set up, we have created the create agent action group response. This is a variable create agent action group response and we have stored the response in this variable. So let's look at the output. Create [Music] agent action group response. So as you can see we have executed the request and we have our response. So this includes an action group ID. It has the action group state which is enabled and before we proceed further we need to ensure that the action group is fully enabled. So although it's already enabled as you can see in this case we'll add a helper function to wait for this state. So this will ensure that you know if you run the entire script from top to bottom everything works smoothly without any timing issues. So let's add that uh helper code. Let's do that. So first we need to extract the action group ID. That's what we have done here. We have to extract the action group ID from the response which will allow us to reference it later. Next we'll give our helper function like how we have given helper functions before. As you can see the target status that we want is enabled. So now that we have made a change to the agent, we also need to prepare the agent again to apply the updates and update the alias so it reflects the latest version. So let's do that. Again, we'll be using helper functions for that. It's enabled now. We you know want it in the prepared state. Agent is prepared. And next is we have to update the alias. So let's do that. So the agent alias has also reached the prepared state. So basically we have as you can see here we have used the the prepare agent function. So now the alias is fully updated and the agent is ready to use. So now you know we can start performing tasks and we can just test it by giving a message. So again first you know for giving the message we are going to reset the session ID. So let's reset the session ID and let's give a different message including an email address. So my name is Nia and we'll just give a dummy name ID like at leia.com. My laptop has display issues. and I want a refund. So this time I've also included as you can see the email address. Now the thing is with this email the agent can confidently match it to the customer ID. So in this case it will just be a random number of course. So this helps simulate a real customer support interaction. So now let's run and see what happens. So first I'll give a helper function and call invoke agent to send the message and put in the response. So firstly we'll just write the helper function agent and print. You pass on the agent ID, agent ID. Just copy paste it because you know we have already used this function before. And right now we have kept the enabled trees as false. So as you can see it has given the chat the user query and the agent response. So it has said hello I'm sorry to hear you have issues with your laptop display. So to better assist you could you please provide some more details about the specific problem you're experiencing. So right now you know as we gave the uh you know email it has already it might have already got the customer ID. So now it's asking about more issues like what's the specific problem you are finding and what is the you know purchase date and everything and also at the end it has asked that please let me know and I'll help you to escalate this to our customer support team if needed. So this is what our agent has responded with. Now let's just keep the enabled trace is true and see what happens. So as you can see here you can you know see the user query the final response and the thing is you can also see the agent's thought process. So this is basically thought process that's going on with the agent which means okay let me give you the details we have so far. It has taken the customer name, customer email, the issue, the request. Issue is that the laptop has displays issues. The request is that you know we want to refund and it has process that it still misses some key details to fully process the refund such as order number or purchase date of the laptop. So without these GCS you know I won't be able to prop this to support team. So let me ask the user about it and that's what it has done over here. So as you can see so far we have built an agent that responds effectively and uses tools to process customer support cases. It can retrieve a customer ID and log a support request making it much more functional. This chapter we'll be seeing what are guard rails in AWS bedrock. Hello also in this chapter we will talk about what are guardrails in Amazon bedrock. So what are guardrails? So basically guardrails in Amazon bedrock are a set of safety measures that help ensure AI generated responses are safe, ethical and appropriate act as filters and restrictions that prevent the AI agent from generating harmful biased or restricted content. So guardrails are essential for regulating AI behavior especially in customer support, healthcare, finance and other sensitive domains where AI generated responses must adhere to compliance standards. So now let's see how do guardrails work. So when an AI model generates a response, guardrails analyze the output before sending it to the user. If the response contains restricted, offensive or policy violating content, the guardrails block or modify the response. So the guardrails in Amazon web can be customized based on business requirements. So users can define content filters to block inappropriate language, deny topics that the AI cannot discuss and harmful category filters to prevent responses related to violence, hate, speech, self harm or explicit content. So these itself are some of the key features of guardrails. In content filters means it can filter and scan AI responses for profanti hate speech or offensive content and block them before they reach the user. In denied topics, predefined topics that the AI is not allowed to discuss. Example political opinions, personal health advice, financial investment recommendations, etc. And then comes harmful categories uh where adjustable safety settings for restricting responses related to hate, speech, self harm, violence, harassment and adult content. And then it has custom word filtering where users can define a list of restricted words or phrases that the AI should not generate in its responses. And then we have custom guardrails via AWS Lambda for like for advanced use cases. Custom logic can be implemented by using AWS Lambda to validate and modify AI responses directly. So now let's see how you know we can implement guardrails in Amazon bedrock with some code examples. So there are basically four steps. So the first step is to create a guardrail. So this is the code for that. So to create a guardrail in Amazon bedrock you have to use the AWS SDK which is 1 2 3 4 Python to define the content filtering restricted topics and sensitivity levels. So basically this code creates a guardrail named customer support guardrail as you can see and it filters harmful language. It blocks specific words example fraud, scam, illegal. It restricts discussions on specific topics like politics, religion and investments and it applies different safety levels to content categories. Now the next step is to associate guardrail with an AI agent. So once a guardrail is created, it needs to be attached to an Amazon web agent to enforce content filtering in AI responses. So this step ensures that the AI agent follows the guardrail rules and does not generate restricted content. And next we have testing guardrails by invoking the AI agent. So we can test if the guardrail works by sending a sample query to the agent and check its response. So as you can see in this code, if the guardrail is applied correctly, the AI agent will refuse to provide investment advice and respond with a safe and predefined answer such as uh you know, I'm sorry, but I cannot provide financial investment advice. And finally, we have step four, which is fine-tuning guardrails. So if the AI agent still provides undesired responses, you can adjust the guardrail by modifying the sensitivity levels or adding new denied topics. So in this code the AI agent will not respond to financial or medical related queries ensuring compliance with regulations. So now why are guardrails so important? Number one to prevent AI misuse. So guardrails restrict AI generated responses preventing harmful biased or unethical replies. Next is to ensure regulatory compliance. So industries like finance, healthcare and legal services must comply with strict rules. So guardrails help prevent AI from providing unauthorized advice and then to enhance customer trust. So well configured AI assistant maintains brand integrity by ensuring all the responses aligned with company policies and ethical guideline and finally to reduce AI risks. So guardrails act as a safety net reducing the risk of AI generated misinformations, hate speech or offensive content. So basically guardrails in Amazon bedrock are a critical feature for ensuring AI models generate safe, ethical and policy complent responses. By implementing content filters, denied topics and harmful category restrictions, businesses can control AI generated output and protect user interactions. And by using AWS Lambda and customizable filters, developers can further fine-tune guard rails based on business needs. Whether it's preventing AI from giving financial advisor, blocking offensive language or ensuring customer support consistency, guardrails help maintain AI reliability and trustworthiness in real world applications. So that was it about guardrails. Now in the next chapter, we'll take it a step further by making the agent smarter. So instead of immediately forwarding request to support, it will perform calculations before sending them. So let's get into it and enhance its decision making. All welcome to the next chapter where we are going to give our agent a code interpreter. So this means the agent will be able to perform accurate calculations by writing and running temporary Python code to support its responses. So let's dive in and see how it works. And before that as we mentioned in all the chapters let's run the code to set up our environment. So let's do that first. Let's set up the environment. So first I'll load all the necessary libraries. Same as before, we move the bottle three ID and we need to input the helper functions. So once that's done, the next step is to create a client, a bedrock agent client, same as before. So let's so let's do that. done. So right now the agent in our environment is same that we used in the last chapter. So to enhance its capabilities, we'll update it and integrate the Amazon bedrock agent with new features. So let's write the code. So let's see what this code is. So this code basically updates an existing action group in Amazon petrop by modifying its function schema and adding a new function. So first we have the function called update agent action group. So this function is called to modify the agents action group. So this ensures the agent can execute the necessary actions when handling customer support requests. And then we have set up the code par parameters. As you can see here we have the action group name which is customer support action. So this names the action group. We have kept it same as before. The action group state is enabled. So since we are modifying the action group we explicitly set its state to enabled. So by default an action group is enabled upon creation. But when updating we must manually specify the state to keep it active. And then we have given the action ID group the agent ID and as you can see the agent version is in draft. So this update applies to the draft version of the agent because bedrock allows multiple agent versions. So changes are first applied in the draft state before being finalized. After that we have the lambda function a. So the agent will execute actions using a lambda function. So the lambda function arn is stored in the lambda function arn variable. Now after that we have defined the function schema. So inside the function as you can see inside the function we have three other functions. The first function is customer ID function. So this function helps the agent identify a customer based on the available details. So as you can see uh this is the description just given get a customer ID given available details. At least one parameter must be sent to the function. This is private information and must not be given to the user. So what this means is that the agent can look up a customer ID but must not share it with the user. So at least one parameter is required. It can be email. These are the parameters. It can be email name or phone number. And then you have the send to support. So if the agent cannot resolve the issue, it can pass a request to your human agent. And then we have a new function an updated function which is called purchase search. So here they have given the description that search for and get details of a purchase made. Details can be used for raising support requests. You can confirm you have this data like for example I found the purchase or you know I can't find the purchase so and so. So basically the agent can search for purchases and confirm whether it found the record. So however private information like order numbers must not be shared. And as you can see these are the uh parameters given over here. The first parameter is the customer ID. The second parameter is the product description. And the third parameter is the purchase date. As you can see the purchase date, the date of purchase should be in this format year, month and date format. So basically this function helps the agent verify past purchases before escalating the case. You know you know before escalating the case the agent must verify that you know if the query is real or not. So for that this helps. So now once everything is set we have stored it to this variable which is update agent action group response. So basically running this agent ensures that it retains previous functions adds the new function. As you can see uh this customer id and send to support function we had already used in the previous chapter. So the new function that we have added is a purchase search function. So let's run this. Now next let's extract the action group ID from the update response. So let's extract the action group ID. So here the variable update agent action group response contains a response from the update agent action group function which we used before. So the response has a nested structure where the action group id is inside the action group. So this line retrieves the action group ID from the response and stores it in a action group ID which is for later use. So action group is enabled. Action group status is enabled. So now let's give it a different message. You can give something like first we'll give the email and Google say I bought a laptop five weeks back And now it has a display issue. I want a sorry refund. So as you can see the message has mentioned the date in weeks. So here we have already mentioned in the function that we need it in this format year, month and date format. But right now we have given it as 5 weeks back. So as you can see we have to do something. So we got to update our agent. So next we have to add a code interpreter to deal with date. So let's do that. Okay. So this code enables the agent to execute Python code dynamically by creating a code interpreter action group. As you can see this is the code interpreter action group. So the process involves three key steps. The first one is creating the code interpreter action group. So the function over here create agent action group is called to create a new action group named code interpreter action. So an action group defines a set of tasks that an agent can perform. And in this case, it enables the agent to execute Python code using the Amazon vet code interpreter feature. So now let's just move on to some key parameters over here. As you can see again we have the action group name, we have the action group status, agent ID, agent version which is still in draft and then we have the parent action group signature which is Amazon code interpreter. So basically this links this action group to Amazon Bedrock's pre-built code interpreter. This enables the agent to execute Python code directly supporting use cases such as calculations, data analysis and automation. So why this is important is because Amazon bread provides built-in action groups like this like for example Amazon code interpreter to handle common AI tasks. So by referencing Amazon.code code interpreter. The agent gains the ability to process and run Python code without needing custom lambda functions and after that we are extracting the action group ID over here. So this ID uniquely identifies a newly created code interpreter action group and it is stored in the code interpreter action group ID for later use. And then we have given a helper function which is used to pause execution until the action group is fully ready. So the next step is to prepare agent and alias to add new action group. So let's do that. So first let's prepare the agent. prepar agent response agent agent 30 Now let's just you know copy the function the target state is prepared and let's run this. Okay, there is some error over here. Let me just fix it. As you can see it is waiting for first it waited for status of prepared then finally it has reached the prepared status. Now once that's done our next step is to update the alias. So let's do that rock update agent alias. We are going to update the image agent ID is equal to agent ID. We have to give the agent als ID and we'll keep the agent alias name as test. And once that's done, wait, let me just command it. Yeah, once that's done, let's just paste our helper function. And let's run the code. So, as you can see, our agent alias has also been updated and reached the prepared state. So now let's set a session ID and write a message again. So first let's set the session ID. Session ID. Yes. And we will write our message. Oh, where was our message? Yeah, this is our message. So, let's write this. Let's run it. Now, let's write a code which prints its response. I'll write the code first and then get into detail. Invoke agent and print. Again here we have the agent ID, agent alies ID, input text which is a message and the session ID and here we have given enable trace as true. So we run it. So basically the invoke agent and print function it sends a message to the agent and retrieves a response. Again here we have kept the enable trees as true. So this is the response. So we have the agent's thought process invocation input again agents thought process. And finally here we have the final response which has said that I have reviewed your request and found that you purchased your laptop approximately 5 weeks ago. Since this is within the typical refund period, I have routed your request for a refund to our support team. So they will review the details and get back to you shortly. You can also reference support case. They have also given an ID sort of thing for you know our ticket or our case. So as you can see what it has done is let's see the agent sort process. So the agent thought process was that the purchase search tool found a purchase for this customer that matches the details provided. So now I need to calculate the purchase date to determine if it is within the refund period. So first to know that because we haven't given it the correct date. We have given it 5 weeks ago. So to determine if it is within the refund period, it has to you know like uh come up with the correct date. So what it has done is it has actually used the Python to calculate the purchase date from 5 weeks ago and basically it has calculated the date over here and then it has come up with the conclusion the final response that it is within the typical refund period. Now let's look at an other code with an another session ID. So let's we'll give it the same message. There is message here. Next, let's set up the client. And basically now we want to retrieve the response in streaming format. So for that first we have to invoke agent response. We have that over here. And next we have to create the event stream. So event stream invoke agent response completion. So for event in event stream if the trunk is in the event. Print this else then this. So this kind of looks cluttered. Let me just format it a bit. So as you can see what happens here is you already know the invoke agent response and the invoke agent function here in this particular part of the code the loop iterates over the event stream and checks for chunk responses. So if the event contains chunk like for event contains chunk the response data is stored in bytes. As you can see in this part, the response data is stored in bytes format. So it must be decoded into a readable string. And the decoded text is formatted as JSON and printed. And for other event types, any metadata or function related events are printed as it is for analysis. So tracing provides insights into how the agent process the request and the conversation remains open allowing continued interaction. So now let's move on to the output. As you can see, this is a pretty big output. So you can take time and see what it is. I will just uh give you a gist about it. Sorry. Yeah, I'll just give you a gist about it. So basically the output represents a trace log of the Amazon bedrock agents orchestration process while handling this case. So as you can see you know it has all the history like the agent has been acting as a frontline customer support service and you know it has taken request from a customer name in India who reported a display issue with the laptop 5 weeks ago. So basically the trace shows the agent step-by-step reasoning and its function calls beginning with retrieving customer ID using the provided email then searching for purchase details using purchase search function which we used and finally escalating the issue to the support team with the send to support function that we used. Uh so the AIS to strict guidelines ensuring accuracy and compliance by leveraging the Python RPL for date calculations and enforcing guard rules to prevent incorrect assumptions. So it also includes model invocation codes, function calls, API responses and trace logs confirming that the request was successfully processed and as you can see finally the agent has also given a final response like I have reviewed Leia's purchase details and routed her refund request to our customer support team. The support team will be in touch with Leia shortly to assist her further with the display issue so and so and please let me know if you need any further assistance. So as you can see the last portion of the output confirms of the response has been finalized and delivered ensuring smooth AIdriven automation customer support workflows. That was it about this chapter. Now the next chapter we will see how to connect this agent to a repository of customer support documents. Hello all, welcome to the next chapter. So in this chapter we will see how to connect the agent to a repository of customer support documents that discusses simple issues the agent can resolve. So once it is connected the agent can resolve some issues directly and redirect others to human workflow. So let's start by importing the necessary basic codes and imports. So the first step of course is to run the code to set up the environment. So let's do that. Next, let's import three. Here we also need to import JSON. And let's import the help functions. Next let's create the client location is equal to 123 dotclient. The service name is bedrop agent region name we'll keep it as US west two. Next let's write the code to retrieve details about the existing agent in Amazon bedrock. So for that describe agent response is equal to the drop agent dot get agent inside we'll give the agent ID So basically the function get agent is used to fetch metadata and configuration details about the specific bedrock agent. So now let's run the code. Then let's print the response. So print JSON describe agent response agent is equal to default here that's string. So this is the response. So basically this code prints the agent details retrieved from the Amazon bedrock in a structured and human readable format. The function JSON dumps over here is used to convert the dictionary stored in describe agent response into a JSON formatted string. So since API responses often return nested data structures, formatting them properly helps in analyzing and debugging the output basically. So the ident is equal to four parameter ensures uh that the JSON output is visually structured with proper spacing making it easier to read. So without identation the response would be printed as a single long line which is difficult to parse manually. And additionally default is equal to str is included to handle any non serializable objects such as timestamps or special data types. So by converting them into strings before printing. So this prevents errors that may occur if the response contains complex data structures that cannot be directly serialized into JSON. So when executed, this command provides a clear view of the agent's configuration including its name as you can see including its name, its status, version, associated action groups and execution settings. So this output is valuable for verifying uh whether the agent is correctly set up checking its state before making modifications or troubleshooting issues related to it configuration. So by displaying all relevant details in an organized manner this step simplifies debugging and ensures that the agent is functioning as expected. But as you can see here they have all the informations including the instructions. So what we can do is let's take that and print it separately. So for that we'll write another piece of code which is print describe agent response agent instruction. instruction. So as you can see this line of code retrieves and prints the agents instruction from the API response. So the describe agent response dictionary contains various metadata about the agent and within it the agent key holds the main details. The instructions field over here specifically stores the system prompt or predefined guidance that tells uh the agent how to behave and respond to queries. So as you can see it has an instruction first. It has an instruction over here. It has just printed a instruction from the uh previous output which is you are a frontline customer support agent for a company and your role is to process customer messages and so and so and also it has uh given guidelines for processing customer messages. So first you have to analyze the customer's message to understand their issue. Then you have to determine the appropriate action. You have to use appropriate tools. You have to provide a summary a clear and concise summary of analysis. And then again for handling issues based on complexity there are two uh types like if it's complex issues or those requiring human intervention or then there are simple issues or general product usage questions. So this is just the instruction printed. So if you go through the output you'll realize there is lack of knowledge base. Okay. So like they have given the instruction they have given the guidelines but still there's a lack of knowledge base. So uh our next step is to add the knowledge base. So let's do that. Let's write that piece of code. So get knowledge. This response agent Get knowledge base sorry knowledge base ID is equal to knowledge base ID. So as you can see this code retrieves information about a knowledge base linked to the Amazon bedrock agent. So knowledge base is basically a collection of stored informations like FAQs or company guidelines that the agent can use to give better responses. So this function over here get knowledge base of is called with knowledge base ID which ensures that the correct knowledge base is assessed. The response is then stored in the get knowledge based response which contains details such as knowledge based name status description and stored documents. So running this function helps check if the knowledge bit exists is it active and properly linked to the agent. So this is useful for troubleshooting cases where the agent is supposed to retrieve information but isn't responding correctly. So now let's just print this and see print JSON dumps get knowledge base response this. So as you can see we have printed and the response contains a lot of configuration. Now next let's connect this knowledge base with our agent. So to do that first let's type the code and then we'll see how it works associate agent. This response is equal to agent. This will say agent base. I'm going to provide the agent ID, the knowledge base ID. The agent version which is drafted And the description which is my knowledge base we'll give it as my KV just format this a bit. So as you can see this code links a knowledge base to Amazon bedrock agent allowing the agent to use stored information when generating responses. So the function over here which is associate agent knowledge base connects a specified knowledge base with knowledgeb ID to the given agent ID and here the agent version is draft which ensures that the change applies to the agents draft version before being finalized and the description we had given it as my KB which provides a short reference name for the knowledgebased association. So basically this function allows the agent to access and reference information from the knowledge base when answering queries and by running this you ensure that the agent can retrieve stored knowledge improving accuracy and response quality and it is useful for customer support FAQs and structured document retrieval. Now let's just look at the output. Associate agent knowledge base response. So as you can see the knowledge base is enabled and we are ready to go. Now next let's prepare the agent and alias. So I will just you know take the code and paste it because we have done this multiple times by now. Now let's see the code. So basically this code prepares the agent for use, updates its areas and ensures everything is fully activated before proceeding. So it consists of four key steps. I'll explain it to you again. So the first function which is prepare agent is called with agent ID instructing Amazon Bedrock to process and activate the latest changes made to the agent such as updated actions on a newly linked knowledge base. waiting for the agent to be ready. So since the preparation process takes time, wait for agent status function ensures that the agent reaches a prepared state before moving forward. And this prevents errors from calling an agent that isn't fully activated. And the next one is updating the agent alias. So since the preparation takes time, so the function update agent alias assigns a new alias like my agent alias to the agent. So an alias is a version reference that allows different agent versions to be accessed in a structured way and this step ensures that the alias is correctly linked to the updated agent. And finally waiting for the alias activation. So wait for agent alias status function ensures that the alias is fully prepared before use and this guarantees that any request made to the agent will work correctly. So these steps finalize all updates and ensures that the agent is fully functional before invocation. And this approach prevents errors, incomplete updates or unresponsive behavior making the agent reliable and ready to process queries. So let's just run this code. Let's wait for it to process. So once it's executed, you can see that we have a fully built customer support agent by now. So now let's just ask it some questions to see how it works. So first let's give the session ID and our message. Again we have given this multiple times by now. created a session ID message is via via.com. I bought a laptop 5 weeks ago and now it has display issues. I want refund. Now let's give the invoke agent function to print. Here we have disabled the trace. Let's print it. So it has actually given us a response saying that I have reviewed a request for a refund on the laptop you purchased about 5 weeks ago and that is now having display issues. So I was able to look up the details of your purchase using the information. Based on my research, I've escalated your case to a customer support team for assistance. So support agent will get in touch. So you can you know continue asking more questions like for example let's try an other message which says let's ask I am unable to connect my laptop Wi-Fi. What can I do? And let's print this. So we have enabled trace. So as you can see this is the output and the guardra guardrail sorry guardrail trace is none. It has given the agent thought process like okay let me see how I can assist you with the issue of not being able to connect your laptop to Wi-Fi and then it has gone through the knowledge base and then it has come up with a final response like to troubleshoot the issue of not being able to connect your laptop to Wi-Fi. There are some steps you can try and it is given to two uh yeah five steps. So similarly you know you can prompt more questions and see the response how the agent responds to your queries. So basically the agent is now designed to follow the conversation flow and use agentic reasoning to decide how to handle each customer request and it doesn't just respond blindly as you have seen. It analyzes the situation and figures out the best course of action. So what makes this agent powerful is its ability to take different paths depending on the problem. So if it's a simple issue, it can resolve it on its own. But when a situation requires human assistance, it knows when to escalate to a support agent. So the way it makes these decisions come down to three things. Number one, the agentic workflow that guides it decision making. Number two, the instructions we provide to shape its behavior. And the third one, the tools and actions it has access to like knowledge bases and APIs that you have seen in this example. So together these elements allow the agent to be efficient, adaptable and capable of handling a variety of customer interactions smoothly. So that was it. In the next chapter I will give you a quick walk through of the Amazon Bedrock console. So now that we have successfully built a customer support agent using code, it's important to explore how the Amazon Bedrock console provides a visual and interactive way to configure AI models, manage agents, and fine-tune various functionalities such as guard rails and knowledge bases. So while coding provides greater flexibility and control, the console simplifies the process by offering an intuitive interface that allows for quick modifications, testing and troubleshooting. So Amazon Bedrock is a part of AWS serverless AI infrastructure, which means it can integrate seamlessly with other AWS services while giving users access to pre-trained AI models from multiple providers. So this makes it an ideal tool for companies and developers looking to build and deploy AI powered solutions without the need for extensive machine learning expertise. So to access Amazon's bedrock you first need to log into your AWS account and once logged in you have to navigate to the search bar and search for bedrock and select Amazon bedrock from the search results. So as you can see on the left hand menu you'll see several options that allow you to explore the different components of bedrock. So these options include model providers, agent management, guard rails, knowledge bases and playgrounds for testing AI models interactively. So by providing a centralized dashboard, the bedrock console allows users to easily manage AIdriven workflows, configure the AI assistance behavior and enhance their knowledge retrieval capabilities. So one of the most important aspects of Amazon Bedrock is a variety of AI models it provides access to. So, Bedrock allows users to choose from models built by leading AI research organizations uh offering a range of capabilities such as text generation, reasoning, summarization and question answering. So, as you can see inside the console, the provider section lists the various AI models available in Amazon web. So, some of these include the Amazon AI, Anttopic, cloud models, AI21 labs, coher meta, mistral AI, stability, etc. So as you can see each of these providers offers different types of uh large language models which can be used depending on the specific requirements of a project. Like for example anthropics claude models are known for the safety first approach while Meta's llama models are designed for high efficiency reasoning tasks. So if you encounter errors when you're trying to use these certain models, it is likely because they have not been enabled for your AWS region or account. So the bedrock console allows users to manually enable access to these models by navigating to the model access settings. So here users can agree to end user license agreements to enable specific models. Modify access permissions to allow Bedrock to integrate with those models and select or deselect models based on their respective project needs. So once a model is enabled, it becomes available for use within agents, knowledge bases and other AIdriven applications. So first let's have a look at agents for configuring and managing AI assistance. So one of the most powerful features of Amazon Bedrock is its ability to create AI powered agents that can interact with users, process queries and provide intelligent responses based on available data. So the agents section in the console provides tools for managing the assistance. So when you open the agent section, you will see a list of existing agents uh along with their current status. So these agents can be in different states as we have already seen while coding. So it can be in the prepared state which means it is ready to use and actively responding to queries and it can also be in not prepared state which means that the agent has been created but it isn't fully uh configured yet. So to create a new agent users can click the create agent enter a name. You'll be having option to enter a name a description and also select the large language model that will power the agent's response. So this step is crucial as the model selection determines the quality, response style and processing capabilities of the agent. So basically the agent builder pro structured interface where users can define detailed agent instructions, choose best suited AI model, enable the code interpreter uh which allows the agent to execute Python scripts when needed and customize its workflow and processing logic to suit business requirements. Now a major part of setting up an agent involves configuring action groups which define what the agent can actually do. So action groups allow the agent to connect with external tools, databases and services to perform tasks beyond simple conversation. So within the console, users can link agents to external services using AWS Lambda functions. Amazon Bedrock simplifies the process by providing a quick create lambda function option which automatically generates sample lambda code, configures the necessary AWS permissions for execution and allows developers to customize Lambda functions for business specific use cases. So for example, if a support agent needs to retrieve a customer order details, an action group can be set up to query a database through a lambda function. So once the function is triggered, the response is processed by agent and delivered back to the user. And also after configuring an agent the console provides an interactive test environment allowing users to send sample queries, debug unexpected agent behavior and ensure proper integration with external systems. Now next over here we have the guardrails for controlling and filtering AI responses. So AI safety and comprehens are critical when deploying automated assistance. So, Amazon Bedrock offers guard rays which we already saw while coding in a previous chapter which allows users to filter and control an agent's responses to ensure it behaves within set guidelines. So, as you can see, unlike action groups which define what an agent can do, guard rules define what it shouldn't do. So, these settings block inappropriate, harmful, or restricted content from being generated. So the console makes it easy to set up and manage guardrails through an intuitive interface allowing users to specify denied topics that the agent should avoid. Define content filters to block inappropriate responses and adjust AI safety levels using interactive sliders. So for example, if an organization wants to prevent an AI assistant from providing financial advice, a denied topic can be configured within the guardrain settings. And then there is knowledge basis which is for expanding the agent's knowledge. So basically knowledge base enhances an AI agent's ability to provide accurate and informed responses by allowing it to access structured data. So this is particularly useful for customer support, internal documentation and FAQ systems. So the console simplifies uh knowledgebased creation by guiding users through entering a name and description for the knowledge base, setting up a access permissions and choosing a data source such as uh you know Amazon S3, Salesforce, SharePoint or a web crawler. So once the knowledge base is set up uh the data needs to be indexed in a vector base to enable fast and efficient retrievals. Amazon bedrock supports multiple vector storage options uh like Amazon open search serverless Aurora Postgra SQL MongoDB Elas and Pine cone. So when users queries the agent the system retrieves relevant information from the knowledge base and integrates it into the agents response. So the Amazon bedrock console provides a comprehensive interface for managing AI agents, configuring workflows and integrating data sources. While coding gives developers full control over implementation, the console simplifies deployment, testing and fine-tuning making it accessible for both technical as well as nontechnical users. So by using all these AI model customization, guardrails for safety and knowledge basis for enhanced understanding, bedrock ensures that AI powered solutions remain accurate, efficient and aligned with the current business needs. So that was it. Congratulations on completing the introduction to agentic workflow course from simply learn. Throughout this course, we explored key concepts of agentic workflows and how they can enhance automation. We learned how to create an agent using Amazon Bedrock connected with the CRM to perform calculations and read an FAQ manual for different information retrieval. We also walk through the Amazon Bedrock console giving you a hands-on understanding of its features. So with these foundational skills, you are now equipped to build intelligent automated workloads that streamline business processes. So we will be making use of couple of tools that are mentioned on the screen like claw chat GPD Gemini sub API and N8N. Right now let's get started with sample how exactly it looks like. Now we are on one of the tools called VAP where we can generate an AI model which will actually interact with you. So by default it has a system prompt related to medical purpose. Now I have a document over here which will explain something to the tool about the identity and how it should react. Right? So let me take up the name of the AI agent Hazel and drop it here. Thank you for calling AI or travel passenger support. How can I help you? Right. And uh here I have the entire prompt. So you might be wondering how did I write this whole prompt. So don't worry about it. You can make use of claude. Let me do that for you. So if you just log into claude and type it as create new and AI travel agent include personality details to be collected name from to take travel like a simple prompt and it's enough it'll automatically create a prompt for you. So there you go your prompt is almost ready. So similarly I have created my prompt over here and now I'll be placing that prompt here and we can also select the openi provider model and uh in case of uh you know selective knowledge base that you wanted to give to your assistant for example if you're working with a specific uh company or flight provider like uh you know Emirates or something like that then you have your personalized prompt or knowledge base and based on that you can provide But right now we will be using knowledge base from open source. So that will be our knowledge source and it will search the web pages for information and then provide your results. Now here you have the transcriber voice configuration and additional configurations tools etc. Here you can add the tools that we will be using. Right? So it's just a demo here how exactly your uh assistant works. So we will go into detail of that. For now just try to check how the agent talks. Thank you for calling uh passenger support. Uh this is Hazel, your scheduling assistant. How may I help you uh today? Hi Hazel, I wanted to book a flight from Hyderabad to Boston and uh my dates of travel will be 15th of May and it would be a round trip. I'd like to fly back to Hyderabad from Boston on 30th of May. Can you please help me with the travel plan? Of course, I'd be happy to help you with your travel plans. Let me confirm um a few details to get started. You're flying so from Hyderrobad HYD to Boston Boss on the 15th of May and returning from Boston to Hyderabad on the 30th of May. Is this correct so far? Yes. How many pass? Great. Is this trip for one passenger or will there be others traveling with you? So if so, are there any children? Uh there are no children. Uh it's just me, one adult traveler. Thank you for confirming. Do you have a preferred travel class, uh, economy, business, or first class? Also, are your dates flexible if we find better fairs on nearby days? U, yes, I'd like to go with first class, and I'm really flexible if you have better options. Got it. First class it is, and I'll also check for better fairs on nearby dates. Uh, just a moment while I look up the available flights for you from Hydrobats, Boston on or around May 15th and returning on or around May 30th. Let me pull up the options. Okay, here's what I found. Outbound flights Hydrobad HYD to Boston. One, flight A12 departs May 15th at 3:20 a.m. arrives at 1:45 p.m. One stop. Four, 85. Two, flight 476 departs May 16th at 10:10 a.m. Arrives at 8:30 p.m. non-stop. 5 0 flight 293 departs May 14th at 11:55 p.m. arrives at 11:15 a.m. One stop 7 return flights Austin US Hyderabbad HYD 1 flight 921 departs May 30th at 6 p.m. Yeah, I'd like to go with the last option. Great choice. Let me summarize your itinerary. You'll be flying from Hyderrobad to Boston on flight 293, departing on May 14th at 11:55 p.m. and returning on flight 921 uh from Boston to Hydrobad on May 30th at 6 p.m. Both in first class. Now, I'll need your full name as it appears on your ID and your date of birth to proceed with the booking. Do you have a frequent flyer number you'd like to add as well? Yeah, I don't have a frequent flyer ID, but uh you can note down my name. So, my name is Tony Stark and my email address would be tony.stargmail.com. So, you can process the tickets and share me the details. Thank you, Tony Stark. I've not noted down uh your details. Now, um I'll process your payment. Could you please provide your card number, expiration date, and billing zip code? Once that's done, I'll confirm your booking and send your itinerary to your email. Yeah. Uh my card number would be uh 1 2 3 4 5 6 7 8 9 10 11 and 12. And uh the expiry date will be December 2026. And uh the zip code to send the billing would be 583203. Thank you for providing the payment details. Uh I'll process your booking right now. Just a moment. All set. You're confirmed on flight 213, departing on May 14th at 11:55 p.m. and returning on flight 921 on May 30th at 6 p.m. Both in first class. Your confirmation number is brg3482. I'll send your itinerary to 20.startgmail.com. You can check in online 24 hours before departure. Please arrive at the airport at least two hours before uh for domestic flights or three hours for international flights. Is there anything else I can assist you with today? Yep, that will be all. Uh Hazel, thank you much for the call and have a nice day. You're welcome, Tony. Thank you for choosing Skybridge Airlines. Have a wonderful trip and a great day ahead. Yeah. Goodbye. Goodbye. Safe travels. So there you go. You saw, right? So it was so detailed. It asked me for the destination. It asked me where am I boarding at including the uh airport codes right? For India HYD and for uh US Boston it said BOS and it also asked for card details and email addresses. But now we are not deeply connected to an actual knowledge base. We not actually connected to an email sender and email recipient. And even the phone number, the credit card details, everything was made up. So for now, since it did not have a knowledge base, it just created all the reports and it shared with us. But what if if we created one in real time where it was actually connected to a knowledge base, it was actually connected to a search engine for it, right? The places of visit and uh the hotel bookings and also the uh reservation for hotels. And apart from that, if it actually wanted to pull out the actual uh uh airport codes, right? All those things, if you actually wanted to create that in real time, we might need a couple of more tools and keys to run with. Right? Now, for that, let's switch to u conversational AI. Uh before that, uh there is no difference much of a difference between VP and uh conversational AI. So, you'll have the same approaches here as well. You'll have a model and inside that you'll have a transcriber. You can choose the voice. So we chose the voice of Elliot. You can also go with any of the female or male voices according to your preference. And here comes the tools. So remember I told you about the NA10 tool and the keys that you're trying to function and also some APIs and all those things. So those come under the tools section. you will have to create some customized functions and tools and add the API keys and search engines and go ahead with that and then you'll have a fully functioning u AI assistant which can process your bookings in real time. Now this is analysis and advanced versions and if you have a specific prompt that you would like to add here like understanding the passenger right each and every passenger will not have the same reason to fly. Somebody might be trying to attend a family gathering. Somebody might be traveling for getting medical attention, right? Based on the circumstance or the activity or the reason of travel, the AI agent should motivate the voice. If it is about the holiday and it should be cheerful. If it is about medical attention, then it should be sympathetic. Right? Such things if you can have here, you can add them here. And going to advanced and you can also have some audio recordings. Let's say if you wanted to have a full conversation recording about the booking for safer side, you can also have that. You can enable them here, right? All those things. Now that we have a brief idea on how exactly we are going to proceed, let's get started with conversational AI. So here on my screen we have the 11 labs conversational AI and here we have the complete N810 functional module where we have the web hook and uh details and all those uh airport codes and uh API keys connected. So we will go through each and every one of these stages for better clarity. Don't worry. Now let's get started with the conversation AI here. So we have named it as Alexis. So call Alexis. Hey there from passenger support. How can I help you today? Hi, Alexis. Uh, I need your help with creating a travel plan for me. Okay, I can definitely help you with that. To start, could you please tell me where you'd like to travel to? I'd like to travel to uh Boston. Great. And where will you be departing from? Uh, I'd like to depart from Hyderabad. Okay. And when would you like to leave and return? Uh, I'd like to start on 15th of May and I'd like to return by 30th of May. Got it. And how many people will be traveling? I'd be a solo traveler. I need your help with planning it for one person, one traveler. Understood. Are there any specific activities you'd like to do in Boston? Uh, I'd like to visit a wedding in Boston. Okay. And lastly, could you please provide your email address so I can send you the travel plan? Yeah, please don't note down my uh email address. Uh that would be rabi.ken at the rate simply.net. Okay, I have all the information I need to confirm. You'd like a travel plan for one person from Hyderabad to Boston departing on May 15th and returning on May 30th with the purpose of attending a wedding. Your email is ravi.karan simplylearn.net. Is that correct? Yes, that's correct. Perfect. Please give me a few moments to look into this. Okay. The travel plan has been emailed to you. Is there anything else I can help you with today? Uh, that'll be all. Uh, Lexus, thank you so much. Have a nice day. Happy I could help with your travel plan. There you go. Now, we will see how exactly did we program this AI. Right now if I go to the homepage here we have all the messages that we need. Starting with the agent we have chosen the language to be English. And if you want some additional languages like Chinese, Italian or Spanish, French, right? You can also go ahead with that. And if you want Hindi, you have other options as well. You can choose any one of these if you want and provided for that. Next, you'll have the first message. How would you like to greet your clients? Hey there, I'm Alexis from passenger support. How can I help you today? After that, I have written a simple prompt here. You're Alexis, a travel agent. Your job is to help the caller create travel plan based on the details they provide. The tool that we will be using is N8. Use this tool to send the caller travel details. So, it will generate a travel plan in real time. And the instructions extract the required details from the caller. Always send the details to the N82. So this will be uh the trigger point in the N8 platform. So it will start collecting the data and once it collects the data it only sends details to the N8 tool. After you use the N8 tool say please give me a few minutes or moments to look into this. So that while it transfers the details, N8N gets into the picture and collects all the details and runs some tests and searches on the web and it gets details suitable to the plan according to the traveler and sends them back. If you are forced to speak in between the process, just say thanks to your patient, I'm almost done. Never say there's an issue with the server or something like that. So that it the passenger thinks you're still working on it or taking some time to write down an email. So there and once the travel plan has been emailed to the callup cheerfully let them know you can also make some modifications to this like I made in the previous one when we were working in Papy considering the requirement of the plan I mean what on what basis is the passenger traveling is it because of medical attention or is it because of a celebration or just a holiday right you can also make some modifications to the existing prompt and ahead get ahead with that and here you can choose the LLM which is responding to You can choose with wide variety of options available here. So we have Germany 2.0 flash as the latest one. So I'd like to go with that. And the term temperature. So the temperature is a parameter which controls the hallucination of the u agent. Let's say if you keep it to the most uh you know uh the complete one then it'll give you most random answers like it'll be too u you know kind of the voice modulation would be too extreme right dramatic type of tone which will make the uh passenger uncomfortable. If you keep it to lowest zero or 25 something like that the modulation of the AI agent would sound too robotic and he'll not be comfortable with that either. So, I'd like to go with the default 50% uh temperature. And here you can also limit the tokens. If you're using a paid version of any of the u modules you're using, then you can limit the tokens. Uh let's say you can use only about 10 to 20 tokens per conversation. You can go ahead with that. And after that, here comes the tools. By default, you'll have one tool with the conversational AI end call. So, what this does it it has an ability to end the call after a certain amount of time. let's say 20 30 seconds of uh no activity it'll automatically end the call or it'll automatically end the call once entire process is done since we are involved with working on N8N web hook which is the trigger for this entire process we will have to go ahead and choose the custom tool option so this is my custom tool so if you get into this you can see the configuration details so the name of this tool will be the this is the first uh block which you need to explain about the tool tool. So this is the N8 tool and you'll have a generate description which tells this tool generates a travel plan once the passenger details and all the relevant data is selected and this is the method working on post. So you'll have all the options here. I'd like to go with post and the URL you'll get this from here on in A10. Let me show you how. So this is the web hook. This is the trigger. If you would like to expand this, you can expand and uh just double click on this. And here you'll have the test URL, just go ahead and just click on this and copy. So if you're working on the production URL, all you do is just eliminate the term test which is over here. Just copy this and you just need to paste it right here. So I've pasted mine over here. Simply Ravi123.app, right? This is the one. And you don't have any headers here and path parameters, query parameters. But the most important one is the body parameters. Remember the tool was collecting your name, your destination of where are you boarding from what date are you traveling and what are your roundtrip details that has the written date and what is the uh activity that you're doing, why are you visiting a certain place, right? All those things are collected here and you need to specify that in terms of body parameters. Firstly collect the passenger details and send them to the request. The first one date uh data type is uh identify the number of uh passengers traveling right and the total number of travelers is just one by default and uh if if in case if you have multiple then you can just provide one to that you can give a detailed number of travelers traveling with you and apart from that proceeding ahead you have uh the data type. So you can provide the data type for date right. Uh by default it has string you can provide the return data type and this will collect the uh return dates of the uh passenger and departure date from which date is he starting to leave and activities. What kind of activities is participating in? Is it like a conference? Is it like a wedding or medical attention? The reason for travel basically the destination where is he traveling to? The origin where is it starting from and email address. collect the email address. And in case if you have further um properties to add, you can also add ahead and you can just uh save uh the changes or you can also choose to create tool or add the tool and your tool will be automatically added here. So, so far so good. We have added or created the first uh trigger thing on the N8 platform. So, this was the trigger part first part and next is the set fields detail right. So here you'll have all the details. Uh so these are the headers and if you close the headers you have the body which we have recently included number of travelers return date and departure date activities destination origin email and you can also uh go ahead with these uh departure dates and then you can also change the data types and everything right and here it'll collect all the details Hyderabad Boston 13th of May right and the next would be the airport codes. So here you will have one of the LLMs. You can choose to have u chatbot chat GBT 4.0 or any of the ones available here. You can just create uh or add one of the details and you'll have to add the parameters. So far so good. So the thing is we have already used uh the quot that we had. So it'll revive after a limited amount of time. All you need to do is create a GPD account and uh let me show you. So I've created a GPD account and here I have named my project as Alexis and I'll generate a secret key. So you'll need to add that secret key to the parameters here and just in case if you have any questions or in doubts you have the option called ask assistant which will walk you through the process and you can add them down and uh here you have the location and dates. So this is a code where you'll have all the details of uh the input that you give uh which date are you traveling and what is the origin and what is the departure and on what date. So you'll have a detailed uh setup here. You can add those things and connect this to the LLM thing blank chain right and after that extend this node to the activities part where you'll have all the activities recorded right similarly you'll have to uh add a traiarch thing here so this is the travely search keys that you need to add you just need to copy this and here you need to paste them right here I've already pasted this so it'll understand what exactly you're searching for for example who Leo Messi and what is the topic and what is the search depth basic right so you need to just copy paste it here and you need to add the details so you just need to paste this URL which I did from here and uh the authentication will be generic go for the generic uh authentication head O and uh just name your head account you can have a custom name here and once done with that you can replicate the same two resource as well all you need to do is have uh the SER API search. So you'll have the Google search API results right here. All these will be linked to your description. So now you just have to paste this particular link first into this particular section which I did already. And then you'll have a generic search credential similar to the previous one. And uh here you need to name your engine and Google hotels. So it'll be using Google hotels for recognizing or searching your hotel state. And then you just have to scroll a little bit or you can just do a quick search. There you go. Google hotels API and you'll have all the required details. And uh similarly you can go for the flights as well. You can create um another block and here you'll have the same sub SER API and uh generic credential type and here the engine will be the same and in place of hotels you'll have to place Google flights. Same procedure here you can just search for flights and here you have it Google flights click on that and you'll have all the details required and uh it will also give you the arrival ID departure ID that we just created outbound date return date adults number of passengers and everything and going back to the canvas here you have the email agent again you will connect one of the LMS to write the perfect emails I'm using cloud here and if you're using cloud then you need to create an anthropic account which I have created just here so that you can be able to create one of the u API keys which you need to add here choose the cl that you're using so that it'll help you in writing an email and sending out the subject and if you're using an API for email as well then you need to link your Google Gmail account and the authentication will be to API so I've connected my email here and that will be I send up and lastly the response. So all these details will be collected the departure date, the arrival date, destination, reason for traveling and your hotel details, the flight details, everything will be in terms of response and here where your web hook ends. And lastly, the entire process gets done and you'll receive an actual email in real time. Now if you want this template, you can let us know. We have already created this and we'll share this with you if required. Firstly, let's go through the agenda for today's session. Firstly, we will understand the basic theories of AI agent. Then, how are AI agents different from LLMs? How do AI agents work? The future risks of AI agents, thoughts on AI agents, and the last and most important point of today's session, which is how to build an AI voice agent. Now that I've made myself clear with the agenda, now that I've made myself clear with the agenda, let's go through a quick theory session on this front. Firstly, what exactly is an AI agent? So the answer at its core, an AI agent is a software program designed to perform tasks autonomously. Unlike traditional software that strictly follows predefined rules, AI agents can make decisions based on understanding, interaction, and realtime data. They leverage cuttingedge models like Chargd, Claude, Gemini, etc. Think of it as a digital assistant that doesn't just follow rigid instructions. Let's say try saying check if Tony is available on Friday and book a meeting. Instead of that, you can simply say schedule a meeting with Tony at the earliest convenient time in the next month. Upon this request, what AI agent will do is it will check your calendar for available slots. After you say that command to the AI agent, what it does is it firstly checks your slots in the calendar. Then it moves to Tony's calendar and finds an available slot which matches with yours and then determine the ideal time, send out invitations and all this happens autonomously. So now that we have a brief idea on how does AI agents work, let's understand the difference between AI agents and the LLMs. While AI agents use large language models like GPD4, they go beyond simple text generation. Here is the key difference. LLMs generate responses based on train data but do not actively interact with real world. AI agents, on the other hand, can access external data, plan tasks, and execute actions all in real time. For example, if you ask an LLM about sports match that happened yesterday, it may or may not have an updated knowledge about it. But on the other hand, AI agents can fetch live data from the web databases or APIs to provide an accurate answer on the spot. Also, there is something called hallucinations. You might be wondering if this term is related to humans. But for a change, LLM can also hallucinate. For example, if you ask an unrealistic question, let's say there are different types of apples like Fuji, Macintosh, Grammy, etc. But if I write, tell me about a watermelon apple or orange mango which are completely unrealistic or in a way different fruits altogether. There are chances where an LLM can give you some false and unrealistic information that sounds believable. This is termed as hallucination. And of course, recent LLMs have a scale to measure their hallucinations as well. Now, let's move into the next part where we will understand how do AI agents work. So, AI agents function as a sophisticated problem solvers with four essential capabilities. Firstly, planning. Next is interacting with tools. Third one is memory and knowledge access. And lastly, executing actions. Firstly, let's go through planning. AI agents begin with a goal. Whether it's researching market trends or drafting an email to break the goal into actionable steps following a chain of thought approach to optimize execution. This removes the need for predefined human triggers. Next stage is interacting with tools. Unlike basic AI models, AI agents can interact with tools browsing the internet, query databases or use APIs. This extends their capabilities beyond static knowledge. Now the third stage which is about the memory and knowledge access. AI agents can retrieve and store information using techniques like R A or also known as retrieval augmented generation. This allows them to integrate a company specific data market research or customer support queries to generate up-to-date responses. And now the last stage which is about executing the actions beyond just processing information. AI agents take action writing reports, managing applications or even communicating with other AI agents. This makes them powerful automation tools capable of independently handling complex workflows. Now the next part of today's session which is about the future and risks of AI agents. AI agents represent a new error in automation and efficiency. However, their autonomous nature comes with risks. Without prior safeguards, an AI agent could misinterpret objectives. For example, if instructed to achieve world peace, an unregulated AI might arrive at extreme solutions. Ensuring human oversight, ethical programming, and responsible deployment will be critical to leveraging AI agents safely and efficiently. Now, some final thoughts on AI agents. AI agents are redefining how we work and interact with technology. With their ability to plan, interact with tools, store knowledge, and execute tasks. They are far more advanced than traditional LLMs. And now to the most important point of today's session, which is how to build an AI voice agent. So to get started with building an AI voice assistant, we will be using a tool called Vappy. So here on the website, you can see some use cases. You can go through them and apart from that you can see some important features about Vappy and some blogs, careers, startups and dogs, etc. Now you can also see an existing Vappy on the dashboard where you can just talk to it. You can give it a try. Let's start with it. Give the permissions. Welcome to Vappy. I'm um an assistant that you can talk to like a person. So, how's your day going? Hi, Vappy. Hey there. I'd love to dive into whatever you want to chat about, but before I do, uh could you tell me um how you found out about Vappy? Uh so I was just uh searching on the websites to you know try to build a agentic AI that talks and responds to my emails, calls and uh booking some appointments. Right. So while I was doing that I just came across you and I wanted to give you a try. That's awesome. Sounds like you're on a cool mission to build an AI agent. So you want to create something that talks, responds to emails, calls, and books appointments. That's a solid plan. What specific features are you thinking about? So um I was thinking about uh a medical uh based approach. So I want an AI agent uh who works for a doctor and uh I wanted that agent to receive some calls from the patients, check availability of the doctors and uh kind of get an appointment the earliest appointment and uh also if uh uh if he or she the patient is an existing uh patient or existing client of the doctor then uh no no worries about the paperwork. If he or she is new to the doctor then they might have to provide some details something like that and you know collecting all the details and checking the calendar of the doctor and uh get an appointment and send out an acknowledgement to the patients about the appointment sort of such thing maybe that sounds like a fantastic use case an AI agent managing patient calls checking doctor availability and handling paperwork is super helpful Plus, it can save a ton of time for both patients and doctors. You've got a clear vision. I hope you get started with this. Uh, thanks for that. Thanks for appreciating my idea. So, uh, I just wanted to deep dive and explore your options, configurations, specifications, and I want to know the possibilities. Basically, let's try a practical approach building one. So, that was my uh call. Sounds like a solid plan. I'd love to help you explore the options and configurations for building your AI agent. You can check out the documentation for all the specs and possibilities. If you want a more practical approach, I can help guide you through the setup or uh any specific features you're curious about. What do you want to uh tackle first? So far, uh I have a plan of action. I just wanted to go back to your dashboard and getting started. So I think uh we're good at uh the point right now. I think I would like to end my conversation or trial conversation with you and get started with the dashboard if that's okay with you. Absolutely. That sounds great. You can head to the dashboard and get get started. Uh if you need any help later, just come back and chat with me. Good luck with your project. Sure. Thank you so much. Have a nice day. You can quit the call now. Thanks. You have a great day, too. Catch you later. So that was really brilliant, right? So I just gave a start I mean you know trial approach to getting started with a conversation with VPY and it was really great talking to him. You know how realistic that sounded correct? I almost felt that I'm talking to an actual candidate you know uh the customer care supporter that kind of a candidate who is an real person. So that was brilliant. Now let's get started with building an AI agent for ourselves. So um so we have model transcriber voice tools analysis and advanced. So as we go ahead with stage bystage approach starting with the model transcriber right so we will be going through all of those you can see here the model so we can use any of the AI providers you can go with Google custom LLM if you have anyone or you can go with gro you can go with uh deepseal you have a ton of uh providers here who can provide you with the model and by default you have open EI here and by default you have uh charg 4 cluster And uh you can also provide some specific files if you maybe if you have written a prompt for yourself and or if it's a specifically designed prompt you can just add the prompt over here and you can see the option which says temperature. So if you go about the information on temperature this says the temperature is used to control the randomness of the output. When you set it higher you'll get more random outputs. When you set it lower towards zero the values are more deterministic. So uh they kind of kept it so official basically it's hallucination. So uh if you give uh a minimum temperature number here then uh the voice of the conversation will be most human and you will feel that you're talking to an actual person. If you nullify it or if you give it to zero then it totally sounds like a robotic uh version which doesn't have any kind of emotions or um you know kind of modulations in the tone. And uh if you put it too high then you can expect uh some overexpressive or overenthusiastic tone which sounds like a madeup uh tone maybe kind of hallucinating sound which doesn't seem comfortable or confident enough. So let's keep a minimum 0.5 which makes it sound more human and in senses. So um an optimum temperature needs to be provided and maximum tokens. uh this is the max number of tokens that an assistant will be allowed to generate in each turn of the conversation. So uh each conver the minimum number of tokens. So if you initiate a conversation with the AI agent, it will consume a minimum or a maximum of 250 tokens to understand the uh input or your command and give you a response. So that's about tokens. And if you come into transcriber here, so you'll have some behavior of the transcriber. You can provide a D RAM and you can provide the language will be English and uh the model will be nova 3 and background there dn noising enabled. Right? So uh let's say uh you wanted to have uh some noise for the background. Let's say you wanted to build u uh an AI agent for hospital then and you wanted your AI agent to be so realistic that even if your client is talking to your agent it must never give uh a simple clue that you're not a human or you're just an AI agent then you can go ahead with this background noising and if you enable it it will provide you with some background noise which is relevant to the domain. Let's say if you're going with the hospital sound, you might be having some rushing sounds like an ambulance call or for you know people roaming around something like that which is similar but sensitive and uh you can also provide some key words for it. Maybe the terms the medical terms like the ICU you know the neurosurgeon surgery options um I ICU oxygen if you wanted to provide some key terms related to the medical field you can add it there and you can also provide the voice configurations so you have the voices over here you can go with any of the options if you wanted some Indian professional then you can go with this basically the voices you wanted to go ahead with right so uh maybe you can choose uh uh Hannah or Paige uh white female deeper tone calming and professional you can go with Harry Lily right so many options to go ahead with so the by default we have Elliot here I think Elliot is the same person which or the same AI bot which spoke to us in the beginning so apart from Elliot we can go with any other person let's go with Lily so or yeah let's go with Lily and you can also have uh you in a trial of Lily and proceeding ahead we have some tools over here which you want to go with if you want to select a specific tool you can go ahead and select a specific tool and some predefined functions basically all the uh uh things here right and now we're in the analysis section which is over here and if you kind of uh wanted to collect the data of what exactly is happening with your conversation or you wanted to collect the information from your clients what kind of uh information was conveyed from your client and what kind of information exchange happened with your AI assistant. How did it respond? You can keep a track on that and uh you can also provide uh a timeout which evaluates the success right it it takes about uh 30 seconds to evaluate if it was successful conversation or not and you can also have some advanced settings uh like privacy, voicemail detection, you know, and all those things and uh call timeout settings as well. It can choose uh you know take u some provided amount of time to wait for a response and if there is no response from the other end the tool or the uh agent can automatically choose to end the call right so that is something which you want to keep in mind and if you want to uh provide multiple options let's say select one for a specific language two for a specific language you can also pro provide that or you can also you know since you're building a medical um AI I agent. Uh you can provide one for neurosurgeon, two for heart surgeon, something like that. If you wanted to in you know input that enable that uh input, you can go ahead and do that. And apart from that, you can also enable uh text messaging configurations uh where you can you know so here there's already um an email or a message that has been inputed here. You can provide uh hi this is uh uh did which voice did we choose? which shows Lily, right? So, what you can do is go to advanced settings and uh we Yeah, this is the one you can write a message. Uh this is Lily from ABC uh wellness partners, right? And uh this is your appointment. Have a wonderful day. That's one thing you can do. And uh idle messages as well. Maximum idle messages, maximum idle time out. And this was the flow for the model and uh this was the latency. So this latency will be uh based on the model that you're choosing and the provider that you're choosing. So uh I find openi and charg4 to be the most optimistic one since it's the default one. So I'd like to go with this. No changes on that front. And on the left hand side you can see uh the assistant workflows phone numbers etc. Let's say you provide an option uh if you get to hear a specific keyword transfer the call to me right you can have uh some specific callers or if the person says it's an emergency I need to consult someone a specific set of keepers you can provide those specific set of keyers and divert the call from the agent to yourself you can also have that option here by providing your phone number and tools and files right so here you might be having all those things that you wanted to upload you can do that. So let's go back to the original page and workflow. You can understand how the model is doing and then you can go ahead with the model so far and you can also create multiple models here. So by default we have Riley over here. So we changed Riley to Lily. Let's see what happens if the change is applied or not. We can find it out soon. And uh if you wanted to create another EI assistant then you can also do that. You can just have a couple of agents followed by Riley. And here you can test and uh you can also talk to the assistant. Once you find comfortable with the performance of your AI agent, you can just go ahead and proceed to publish. Now let's write the prompt here. So I've already uh got the prompt here. So this is the prompt. I can just uh copy it and uh paste it in the place of prompt and then I can just proceed and uh try talking to my assistant shortly. Uh, the mic will be activated and I can hear something from the agent. Thank you for calling Wellness Partners. This is Riley, your scheduling assistant. How may I help you today? Yeah, you're Riley. Fine. Hi, Riley. Uh, how are you? Hello. I'm doing well. Thank you for asking. How can I assist you with scheduling an appointment today? Uh so uh I wanted to meet uh a neurosurgeon um and I wanted to get some scans done and I wanted to know how my uh brain is doing and proceed with some medications to solve my migraine problems. I'd be happy to help you with scheduling an appointment with a neurosurgeon. Have you visited our clinic before or will this be your first appointment with us? Uh yeah, Riley, this is my first visit to the hospital and uh I don't have any previous visits to your hospital. This would be my first time. Great. Since this is your first visit, I'll need to collect some basic information. Could I have your full name, date of birth, and a phone number where we can reach you? Yeah, my name would be uh Tony Stark and uh you can reach me out at uh99861 234. Thank you Tony and could you please provide your date of birth as well? Uh my date of birth wi