Building Enterprise-Ready Agents using Agent Bricks

Key Takeaways

Things look really good, really optimistic. Every new day there is some new person releasing something really exciting which might suddenly replace all the software developers and whatnot. But it's not exactly the same in enterprise AI. A company that's considering AI governance and privacy and all these things in mind, it also shows the maturity of that company, of the leadership, of the team. How many people here are dealing with AI agents or you know anything related to that at work? Okay, so a good chunk of it. Perfect. My name is Harshett. I am a resident solutions architect at datab bricks. I did my masters in data science from the University of Washington Udub. uh been in this field for about more than eight years and I'm directly working with uh enterprise customers or data bricks enterprise customers in manufacturing and healthcare life sciences sectors which means you know they come up with all these different challenges I have the data I want to do this with my data and then I help them uh you know implement those things we are going to look at state of AI agents uh then we look at then we'll look at an enterprise scenario basically a use case then uh I'll briefly or or in a very detailed way go through agent breaks then I'll give a demo and then key takeaways and we'll have some time for question and answers now the question the first one is state of AI agents and if you were to look at the like what I call the family and friends version which is basically outside of your enterprise work what it means every day you have got you know new models coming in cursor uh I don't know entropic all these things new open-source agents uh uh I don't know how many of you know the open claw cloudbot and whatnot so things look really good really optimistic every new day there is some new person releasing something really exciting which might suddenly replace all the software developers and whatnot but things seem to be working out really well but it's not exact Exactly the same in enterprise AI or enterprise world. And when I say this, this is not really out of uh you know uh some sort of sales pitch. It's something that I do each and every day at my work. So uh one of the things that really uh recently came out is the data bricks publishing a state of AI agents uh report and it provides a snapshot of like how organizations are shaping their uh data and AI strategies and priorities and the source is not the trust me bro benchmark or all those things it's actually the insights that we have collected from you know 20,000 plus customers customers that we have all across the world. So we get to see how they are using the platform, what are the difficulties they are dealing with, how they resolve it, all those things. Some of the summaries uh some of the key takeaways from that uh report are here u and I'll share a link where you can go through all the details of that report. So the first thing is multimodel strate uh strategy is the way to go. Earlier when the when uh open AIS you know and the Geminis and the cursor and chat GPTs of the world came out people started you know uh companies tried to pick one model but then over the time they realized that this is getting too tiring. Every single day there is a new model and that you know claims to beat everything else out there. So the best way to deal with this is you adopt a strategy which allows you to just use multiple models. So 78% of the companies are now using two or more LLM families. The other one is since January of 2025 basically last year so within a one year there has been a 7x jump in investment in AI governance and security. So earlier when things were you know really new really exciting there were a lot of passion around building and shipping these agents but as dust settled down you saw you you see increased concerns around AI governance and everything else and so the other thing is companies that use AI governance were putting almost 20 uh 12 times more projects in production and you might say hey well this is just a correlation it's not causation but the one of the things here is to consider the fact that a company that's considering AI governance and privacy and all these things in mind it also shows the maturity of that company of the leadership of the team right so that's why when all of these different things come together that's why you have 12 times more things in production another thing agents are driving the core database activity ities. So two or 3 years ago there were basically no AI agents or at least the way in the form we know of them today really popped up in 2024 and then rose sharply in 2025. So what it says is you know this this chart simply says that on the platform that data bricks I mean neon that uh one of the platforms that we acquired uh yes 80% of the databases were actually being created by the AI agents and this is just the database when you look at the branches that gets even you know steeper it's 97%. Let's look at the a typical enterprise scenario. So let's say there is a bank called Nova Bank. Uh it is a global financial services firm. It needs to monitor threats and uh that needs to be done in real time. How can that be done? Uh you do it by you know monitoring the authentication logs, endpoints, user access, threat intelligence data so on and so forth. And you and your bestie belong to the data team which is in charge of doing this. So you have got your best friend as well and you're both uh tagging along on this ride. Simply put, your task is to correlate the data or use the correlated data and contain the threat quickly. That's all the business that's all that business cares about. It doesn't matter how you do it. You build dashboards, you build agents, you you know climb the building. Doesn't matter. That's the end goal. So next one would be well that sounds pretty straightforward. We know the problem. We know what the solution should look like. I said you know there will be logs and all these things. So basically you need the structured and unstructured data. You need a bunch of AI agents and then you need to slap on a very fancy nicel looking UI so that if your boss or any manager or a director looks at it they get a nice feeling of it. So, well, now let's quickly build some AI agents, but there is always a but and there's always a meme. There is something called agent sprawl. And what it simply means is you have got lowquality reasoning on enterprise data. When you have these AI agents which are or forget AI agents for the matter when you have LLMs the large language models they are trained on things that are available on the internet right but the data that is sitting in your enterprise that's not really available on the internet. So it has absolutely zero clue of what your data looks like what's the business logic inside of it is no idea then it guesses and you will see what I mean by that. So that's one. Number two, no visibility or auditability. Well, you you decide to, you know, the the best way to deliver some results fastest in the in the best in the fastest manner is to maybe make an API call, send in some data, get the data, get the result, and say, "Hey, I got the result." But then the problem is you have absolutely no idea what's happening behind the scene with with your data. Is the LLM storing it? what kind of data you are sending, what kind of response you are getting. Nothing on that front. Too many AI vendors nowadays. I mean uh I mean yeah you have got Google, you have got anthropic and you have got OpenAI. There are and you have got Meta as well and there are a couple of others and then you have the last problem is you have no way to measure the quality right. So every person has or every team comes up with their own standards of like what a nice agent or what an accurate uh agentic system looks like and but that's what it is at the end of the day a team-based metric of what quality looks like and to sum it up the governance is the nightmare part the the governance and the you know everything that comes with it. And so in case you still needed some convincing, this was a post that I saw I think two or three days ago and I'll just uh you know ask you to read it maybe 10 seconds. Just look at the the box the red box. It says built an AI agent. It was answering some leadership you know it was answering to the leadership about metrics. The leadership made a whole bunch of decisions on it, but then I later found out that it was hallucinating numbers for the past three months and the person at the end says I caught it by accident. So nobody wasn't even like questioning the the outputs that the the model was generating by pure luck. And so this is where something called agent bricks comes in. What is agent bricks? It's that one unified agentic platform that cures the sprawl. And what is cures? It's a it's a word. It's a big word. It's a fancy word, but I'll give you the the the meaning around it or behind it. But before that, let's look at what's happening in in this uh architecture or in this platform. At the base of it all or at the core of it all is what we have uh is is Unity catalog. So, and that's the governance layer. And so, any AI gateway models, MCP agents or EVAs, files, tables, anything and everything related to your data and AI is directly governed by Unity catalog. You don't need to slap on an additional governance layer. On top of Unity catalog, you have something called MLflow that Z covered. And that MLflow is useful in uh really kind of monitoring and observing and evaluating a whole bunch of things, a whole bunch of components that sit at different levels. So that is vertically uh uh it it is uh vertically spread out. This is horizontally spread out. So this means a combination of them gives you a good coverage. on top of Unity catalog and along with MLFlow you'll be building models you'll have you know data serving you'll have uh vector store vector search feature store you have you'll have different libraries to train your model and you'll use different uh you know developer tools to do that on top of that I mean or or using these things you will do a whole bunch of uh these things what it mean uh what will you be doing you'll be doing data reasoning and retrieval using the structure ructured unstructured data, maybe document processing and other things. Then you have got agent factory and this is what I'll be covering in my demo. You will have a supervisor agent which can work with a bunch of other agents uh that live in your data platform and uh it's able to answer questions. Then you have got agent deployment. um you can do lang chain and crew AI and I don't know how to even say DSPI or DSP uh but yeah you have a whole bunch of things here and then you have different model providers and so that's so basically the point that I'm trying to make here is this one platform is what really helps the enterprise team to really not care or or really not spend a whole lot of time building the foundational layers but start using what's already available what's already built out for you and quickly deliver value and so if you remember I covered cure so what is cure I'll briefly cover it so your problem was initially that I shared was we had lowquality reasoning on enterprise data and the solution for that was contextual reasoning and that contextual reasoning is uh coming from your business data which resides in data bricks when you are when you as the user are using the platform. There is some sort of learning that's that happens behind the scene depending on how you use it, what kind of questions you ask and so that's contextual reasoning. The other problem was no governance data leakage and absolutely no visibility. For that you have unity catalog which provides you the unified governance. Next one is you have too many vendors and a and a fragmented stack and a solution for that is the run on any model or run on any framework that comes built in uh with agent bricks. The last one is uh no measurement and and your agent is degrading silently just like the example we saw earlier and the solution for that was uh the evaluate and improve method that comes with MLflow. So this is what a data pipeline would look like. you have logs from at 10 different places. I mean this is not exactly what it look like but this is just a sample of like what a job would look like which brings the data from all these different places. So to represent that I used a block called generate data then you will do a whole bunch of SQL transformations and so on and so forth. Now once you have the data in data bricks it will look something like this in your catalog. So I have a uh a catalog called agent bricks demo. Within that I have a schema called FSI cyber and I have whole bunch of these tables and these tables are have all gone through the medallion layer you know the the transformation. So raw uh sorry bronze, silver and gold. So here you can see raw, silver, gold and everything in between. Now I'll very quickly cover this. One of the things that I myself I have I did hear about it earlier but something that I tried today so I had no plans of covering it but here I am. Once you have the data in your catalog the the the first thing you would say is hey let me see let me visualize what that data or the the final uh massage data or the gold layer data can do for me. So you would start with building some sort of dashboard at least that's what you would do now within data bricks or and and as part of the whole agent bricks platform what you have now is you get to say uh and this is the dashboard section you say add sources you'll go here u select all the let's say gold level tables so we'll not spend a whole lot of time but I'll I just wanted to show this very quickly. Okay, so it has I I added and uh I added a whole bunch of tables that I was interested in visualize visualizing uh on my dashboard. The next step would have been me going to this you uh untitled page and start to add some uh you know visualizations. I'm not going to do that. There is something called this assisted authoring the agent mode and what I will do is I'll simply click on this button called create a sample or simp yeah yeah sample dashboard and then off it goes it will look at my data it will try to make sense of it will find out what it has what kind of insights it can provide and so as you can see it's already starting to think it will do its thing we'll come back to it at the end of the demo but this is what I was able to you know create without typing literally not typing a single word apart from that pressing that button. So it went through a whole bunch of my tables identified the comments or anything that I have added and so as you can see it has added a whole bunch of these things on its own. The other thing you would want to do is you know you have visualized the data. Now what if you want to really talk to your data and so for that we have a tool called Genie. It's again uh a conversational analytics tool which basically lets you talk to your uh data uh or which lets business users talk to you talk to your data and the way it works is you would here go ahead and click new and you will add your data. I'll just add one single thing so I can show what it does. And so once I finish doing that, I have this ready to go. I'll add some more instructions to explain or or to instruct the LLM how to behave, what to do, what not to do. I'll add a couple of these SQL expressions. Add some business logic inside it just to get it ready and just to help it understand my data better. But once I have done that, what I have right now is an interface with which business users can start talking. And what I'm doing is there are some ready-made questions the agent thinks might be most relevant to the users. So I'll just click on it. So and the question was what what recent temporal patterns e emerge in this data set. So basically I'm asking during the whole uh duration for which this data is available do you see any sort of temporal patterns and so here you can see uh on the right hand side when I expand it you can exactly check line by line of what ex uh what it's doing so you see it's looking at the data set time range it identified first from to date ranges then what are the business impacts and whole bunch of other things. I mean uh and and so having this can really help you confirm that your LLM is not really hallucinating, right? And this does not really need a technical person to jump in and look at things. I could be a business user and I don't have to do I I probably won't be seeing this. I'll just have this interface and while this uh agent tries to answer my question I all I have to do is click on this and all of this shows up. So as a business user I can yeah uh there is this something called datab bricks one. So you will add on your uh business users to your datab bricks platform and uh you will shoot out this link and they will have access to it. The clusters will run of course because the data the the agent is gen uh generating the response. It's actively querying the data behind it. So the warehouse the data warehouse will need to run. But yeah the data bricks one thing that's a new thing uh it takes away a lot of complexity. Now while we were while I was yapping as you can see it went through the whole uh data set and it identified these sort of authentication anomalies and patterns and events and all those things and uh you can uh edit the visualization you can do a whole bunch of other things you can download it as a PDF that there's no glory to that but now the other thing that uh the the coming to the main topic you have agent breaks uh this is the uh available here at the bottom corner. Once you click on that you'll see a whole bunch of options or different types of agents that you can build. So one is document parsing information extraction knowledge assistant AI genie. So the thing that I was showing you earlier the AI genie that is part of the agent bricks platform. Now let's say I'll I'll try to finish this uh soon. So there are two things that I'll be covering. Number one is knowledge assistant. Number two is uh multi- aent supervisor. To build a knowledge assistant, what you will do is you'll see this guy over here. You'll click on build and it brings up a very minimalist interface for you where you provide the name, you provide the description, and you say where are my files on which I want this to be built. And so here they are. They can be in the PDF format. So I'll go here and I'll select this PDF uh folder. I can just name it something and in the options I'll say hey you need you are a so and so knowledge assistant you need to act in this certain way and once I click on that well maybe it's asking for agent as well I mean description. So I do some of this and it goes off to create the the knowledge assistant. It will take about you know 10 uh 5 10 minutes to parse through all the data all the PDF document and everything else that you have but in the end you will have something like this. So you have the the knowledge assistant built here and I'll ask it a very uh simple question. So this this knowledge assistant has been fed a whole bunch of PDF files which cover topics around you know uh uh what are my SOC or it has documentation around security policies uh in in the organization and again here you you can see everything that the agent is doing you can you are able to click on that thinking tab and it will you know come back with uh details on what exactly it's doing And so now it's able to give you the answers that you are looking for. For PDF files, it can be in any form. In the PDF for form, you may have a mix of free flowing text. You can have tables and everything in between. Right? So it will be able to pick that up. And if you go to the documentation, it uh lists down all the supported formats that uh you know that will work with this. Okay. But basically uh you you saw this whole bunch of uh response coming down and towards the end it also sites the page or the section from where it got the information. So you have uh for for threat correlation workflow I don't know I have never read this PDF file and probably most of the analyst will probably never will they just want to understand or they want to get an answer and so if you want to kind of refer that or confirm whether the answer is correct you can always click on those citations and get the answers and so that is the the knowledge assistant I'll very quickly cover the multi- aent supervisor because this is the part which kind of gets exciting. The multi- aent supervisor if you remember just few minutes ago we had created genie uh the thing that loves working with structured data and we also created knowledge assistant. Knowledge assistant loves working with unstructured data and so the multi- aent supervisor can really bring these both of them together. it can understand whether the question that you are asking needs to be answered by the uh knowledge assistant or by genie or by combination of both of them. So I'll copy this question. It's not something that I have there is no predefined answer to this. It will go through the the entire stack to figure out where it's getting the data from or where it should get the data from. So if you see here I was asking according to the SOC playbook I asked it a question related to my SOC playbook and it immediately identified that this needs to go through the knowledge assistant right and so then it asked it a question and then it came back with oh this is a bit too fidgety okay so yeah so basically it does a whole bunch of things and uh it gets you the answer you have the knowledge assistant going first then it will come back it will maybe ask questions to genie and so it will get you the answer is is the point that I'm trying to make now this was something that I had asked it so earlier you know I asked it you know what are our am and privilege access policies and so it said okay let me look at that and then whole bunch of things come back and then everything that it mentions there are these citations that you can click on and get access to the last thing that I would proh like to cover is the integration with MLflow. So I built it all I did was you know a a simple user interface. Click through things selected a whole bunch of things and off it went. Now let's say where uh let's see where observability and all these things come into picture. I will go to experiments and this is where uh you'll see the integration with MLflow. So there are a couple of things you have the multi- aent supervisor you have knowledge assistant let's say I want to look at what multi- aent supervisor was doing so here you can look at each trace of like how the whole thing went asked a question what did the model do what was the context like uh then what was the response whole bunch of those things uh I can again go back uh you know it it will probably take several hours for me to for everything but you can also look at each session like how different users interacted with different uh in in their different sessions. You can also create judges meaning LLM based judge where you are constantly monitoring whether the response uh you know given by your LLM or this whole agentic system is compliant with certain rules and regulations that you may have. So let's say safety is one of them. You can create another LLM judge. It will ask you to select a type. So here, so guidelines, relevance to query and a whole bunch of these things. Basically, what did you achieve after this? Just because you have this platform where you don't really have to uh you know worry about the governance, you don't have to worry about observability. Everything comes baked in. What you were able to do is you were able to immediately identify you know the kind of data you have, the kind of insights you can drive out of it. It's a good platform. That's what I would say. The key takeaways uh for me uh and for that I would love everyone else to take as well is agents are now creating most of the databases and the branches and this is all happening to uh you know because of wipe coding the new term wipe coding. Now we can discuss whether wipe coding is good or it just adds to tech debt. Irrespective of that it's happening. A vast majority of enterprises have adopted the multimodal uh multimodel approach. The enterprises that are investing in governance, they are shipping 12 times more AI projects in production. already mentioned. One of the other things is enterprises have many high value use cases for agents but they are mainly limited by these different sprawls that you know the other challenges that I had mentioned and because data bricks had that advantage of being able to see what's happening with their client it gave it a lot clarity about you know what kind of challenges they are dealing with what kind of solutions will be helpful so on and so forth and that would help them help the enterprise cure the agent sprawl. And that's pretty much it. Uh if you have any questions or if you would like to connect with me on LinkedIn, it's here. Thank you.