Your AI Interface is More Important than the Content | Humans Talking Agents Episode 1

Key Takeaways

agent interfaces are more important than agent content I'm Daniel I'm bradden let's jump into it today we're talking about interfaces uh so interface is what you're actually using when you uh use a chatbot voice agent whatever it is it's that end interface for the customer and we believe that it is more important than the content which is a bit of a hot take so Daniel I want to kick this off by saying like what are some of the cool interfaces that you've seen on the interweb yeah so I think like just starting from scratch right um when you look at a lot of like chat interfaces it's traditionally been this like you know chat widget in the bottom right corner um they are like very isolated from the actual website itself and so it's just an experience like on its own and so I think the thing that started to change that was obviously when you had J ke and all of a sudden you had a full kind of like web experience um but where I'm starting to see things going now is moving away from the idea of this agent being like an isolated thing that is separate from your website or your web app uh and it's a lot more embedded and so in actual apps you know you have this concept of like a co-pilot right some that's writing alongside you and actually interacting with the app itself so like curs is a great example um and I think the most interesting thing that I'm starting to see is that um you know co-pilots aren't just for apps um we're starting to see them like a lot of these agents interact more with like a website yeah and so I think where I'm seeing the biggest change in honestly the change that probably appeals most like most people of the mass Market is is how our like website is changing yeah uh and how are agents moving from this thing that is kind of right now isolated from the website to something that's actually more native within the website itself I yeah you're seeing a lot of like different experiments with the interface like you're seeing like opening eyes recent like operators where yeah like the interface is just the the app you're actually using it's not even like a chat interface per se anymore like the interface is just the app itself even saw like coher one of their Founders recently tweeted they're hiring an interface designer to explore new interface types right cuz the chat widget in the bottom right corner has kind of been like the Paradigm for like 10 plus years now but the capabilities have grown by like orders of magnitude but yet for most chatbots it's still like a little bottom right corner thing yeah yeah and I think like there's there's is we separating out two different parts right so one of them is you have like the core base models themselves and I think all of these are moving towards being this kind of you know omnipresent like assistant for you yeah and so I think one question is like okay for the base model like what is the best type of interface for that given that it's supposed to be um or it kind of moving towards being a system for you right to umw the web being able to search the web being able to like present different items to you like Cloud ARS right actually write the code for you do you like Cloud artifacts no I mean it's good for it's good for specific things right I think the other part of this is like certain interfaces are good for specific tasks yeah right like when I'm uh like I need help with writing code like like artifacts is great because it shows me all the code there what I'm trying to like understand like what a front end might look like um like vzer interface is really great as well is rendering the front end right there but when like asking it questions like or like trying to search the web like I don't need the artifact right like sometimes I'll ask it for like up the prompt you're asking a question and they like a write me code and I'm like this is not what I'm asking yeah it's like really amazing when I remember I was doing some work for a customer and we're building like an agent we like uploaded a like PDF Doc and I wanted to create like a mermaid diagram to show all the different flows yeah and then it you know beautifully rendered this like mermaid diagram on the you know on the right like that's awesome and then other times it's like give me like restaurant commendations and it's like generating code you're like too much too much yeah yeah yeah and I think a good example actually is uh like perplex like I think they've nailed their interface for like searching the web so if I go ahead and be like um how many solar panels does it take to power the US um and I'll use it perplexity a lot for like web search because I I think that the way that they created their UI around searching the web uh providing you sources being able to show you like videos that it's coming from um is really effective but you know I don't really go to I don't really go to um perplexity for help with code well you know what's cool about um a lot of the foundation model interfaces are starting to introduce the um almost showing like the Chain of Thought reasoning behind the scenes like deep seek does a good example or does a good job with that the 01 models do a good job with that and it's funny you're actually starting to see that on the like bottom like customer support facing chat Bots you're starting to see people like emulate that type of interface experience where has like thinking reasoning etc etc and you can kind of like spoof it too like sometimes you don't even have to show the Chan of thought like if you're using a a really small model sometimes it's like in the interface itself you'll like be like hey when you're making an API call have it like say like I'm looking that up for you or like if you have the uh you're like looking up a DB record or something being like pulling your information like being able to actually just give some feedback to the user on the interface of what the agent is doing whether or not it's actually that reasoning is coming from the Ln model itself yeah but I think it makes sense in terms of like understanding what's happening behind the scenes right cuz even when you're talking to someone normally like and and I say this to people are building like voice agents all the time on the phone right um your agent is going to do a lot of stuff like look up your your your phone number in a database pull information it's all the same thing as a human would on the other side of the line but the difference is when a human's doing it they're they're like telling you like hey give me a second I'm going to like look up yourself in a database versus just being silent on the other end and so part of it I think is like being able to understand like a lot of these things have latency involved with them and so how do you best show someone what's happening behind the scenes so that they're more understanding of the latency so rather than being just like is taking forever respond it's like oh it's actually doing things behind the scenes well this is what's going on even yeah like with voice interfaces in particular like the whole use of like filler words right yeah like um you know I hear a lot from folks that they want their like phone ivr that's doing lots of API calls and like validations behind the scenes to uh have the same like latency as feels like chat gbt like advanced voice mode yeah but often it's like there's no matter what you do like you cannot make an API call faster it's always going to be like a second to two seconds if you have an internal API maybe like four 5 Seconds right and so being a with some pretty shitty apis yeah that's you know we've we've seen a bunch right um and so being able to use like a filler word where the agent is just like kind of filling the D air a little bit I mean even actually I mean it's kind of like the modern progress bar right like what's a progress bar for agents like that's that's really what this is yeah well I mean we've all used those like voice ivrs that are like super traditional and it's like you know press one press two and yeah those are pretty fast um but there's often just a ton of Dead Space so even seeing Innovations on the The Voice interface of like background audio you're seeing yeah uh audio cues just like letting the user know like hey we heard you there's actually I I saw a really interesting interface where um they had input cues and so as you were speaking to like you know over the phone what the agent would actually do is play like key like keyboard clicking sounds just to let you know that like it's thinking it's actually hearing you and then even then like once you stop speaking the keyboard clicking stops but you know that it was like it was here these things are actually like like Innovative on the interface side right cuz like really all this is especially like and I think maybe putting phone calls aside right because the like what all these voice agents on the phone are trying to get to is just get one to one with a human so someone doesn't know that they're talking to an right because you don't want to talk to an asent but I I don't think that's like actually being Innovative you're just trying to like you know all this background noise and stuff like why is background noise even good right like like I'm on a call I'm trying to go to a quiet area so I think but like moving aside from it fills the space though right like cuz I think want to tell the user like hey the thing didn't break you know what I mean like if you have two seconds of latency you know maybe that's like a standard API you don't want the user especially with bargeon 2 on voice yeah as an interface it's like if you hear two seconds of Dead Space the user might start saying other things which might actually break the bot right it might activate the bar I agree with that but I think that like at least when it comes to the topic of like interfaces I think where the most interesting thought is happening is is on the web um is like visual interfaces right like I think you're very limited with what you can do over like audio only and so that's why when I look at a lot of the stuff that you see with the basic models right with like with perplexity having a specific UI for search with Claude having a UI honestly that's best for like rendering code and like undering front elements and then I think open AI trying to figure out what makes sense for them yeah um is kind of interesting because the other thing I see too is like you know going back to the original argument of you have one angle that's like all of these base models trying to get to some sort of like specialized assistant yeah right and then on the other hand you have um uh people's businesses and their websites yeah right and like you make use your agent to Surf a lot of these websites and return information to you but at the end of the day if you need like a specific service sometimes you're actually going to that website you're going to that business to figure out what they do right and so that's where I'm starting to see more stuff as well where it's like what do business specific agents look like and how do you how does it change the way that people interact with uh companies and their websites well I I think the I think kind of like the transformation we're seeing is when agents first launched it was like you know your chat bot in the bottom right corner and it was like a small little widget on top of your core experience which was like the website and the mobile app or whatever but I think now you're seeing like the so that's like one extreme right where the chat bot is just like a you know like an added-on widget The Other Extreme is perplexity where the whole agent is like that is the app right or chaty BT like the whole experience is the app yeah and so you're starting to see um some hybrids in the Middle where instead of just having the agent be like a little thing in the bottom right it's actually becoming the experience itself and I think it begs a question like do what do websites look like for brands in like 5 years as more companies move along that Spectrum eventually like if you go to like a large banking website is it just an agent is it like your branch experience yeah and I think that's what what that's I think the Insight is like like for me it's websites are are moving towards just becoming web apps yeah um and the experience is moving from uh what you know web initially started as a like digital business card yeah and so I was having a conversation with a couple webflow agencies and um it was kind of interesting to see how they're thinking about it to and one of them was this web flow agency um in Australia and he specializes in conversion rate optimization right and we had a good combo on this where he works a lot with small businesses and for small businesses a lot of it is very transactional like you are coming to the website a plumer example right like you're basically you want to quote and you're calling them because you want to figure out like hey do they do what I need them to do how much is this thing going to cost me can they come on time and initially you know websites were basically just digital cards they just have the number of the website they told you a little bit about it and over time websites started to become closer to actually talking to the end person themselves talking to the end business or service and so what he focused on was you know figuring out how do we um get that experience closer and you know would build things like calculator is like figuring out for decking right what's the size of my deck what materials do I want get me as close as possible to quote and then actually go and talk to the business owner and I think that that idea of like moving from a like virtual business card into something that um is almost as good as talking to someone who's an expert about that service is kind of where a lot of these tools are going and I think like where we're starting to see is is is Agents move out of just chat widget and kind of be more embedded Within These core UI elements yeah and so I'll s a couple examples or at least one that I thought was pretty cool so um someone the community built this uh for an e-commerce website called sha so this company does uh kind of like high-end portable batteries cool right and so this is a really easy example because for e-commerce like a very easy way is like the FAQ I think like an easy thing to do right when you are like buying something expensive like this is the is the FAQ right you have a lot of questions about this like for me if I'm buying this like I I care less about the battery and I care more about like what can it charge and can it charge situation that like I'm looking for and so a lot of times people go to FAQs but what they did that I thought was pretty cool is they kind of invented an agent in the product page to replace like this FAQ section yeah so now instead of trying to scroll through and figure out like will this charge what I need to I can just go ahead and ask it like okay I'm doing an outdoor like outdoor music event um with like two PA speakers um and my MacBook Pro uh can like how how many hours does that left bar does that change the the topic of conversation is that changing like the system Persona this one yeah um so you'll see so actually like what's going to do is it's going to save my search um into this left bar here um after it's gone and done it but here's another example right it's like it looks like he's got the actual uh what's going on in the background too so it's like cool are you going to use these individually or simultaneously I'm going to use them all simultaneously um oh this pretty cool sech internet yeah and so all of a sudden right like I'm now this is the high-end product and the same way that I'd be talking to a sales like a a rep or someone who's an expert I'm now talking to this part of the website um and it's making the actual website less of a static website and more of something that's actually uh that I that I really engage with right and so and it's got it's interesting how the streaming comes in for both the the follow-up question but also the content yeah looks like he's running in parallel to kind of increases the speed um but yeah so it's cool so for me I'm like okay great so this is actually going to run it for like about nine hours like I can probably go and give it more information here um and then it saved my searches on the left hand side here that's really cool and so this is where I'm starting to see um websites move more towards this is kind of that like hybrid I was talking about where like the whole website is not an agent but the agent is doing far more than just answer basic if we actually look back to like the original chat Bots for a lot of these websites when you have a big banking website where it has like let's say a thousand Pages yeah often the chat Bots were actually just better navigation of the website cuz the content was all there you were able to I don't know do like a you know electronic funds transfer you were able to find like the ability to you know activate a new card but there were so many pages that sometimes a natural language interface with just some buttons was easier navigation of the website but it wasn't adding any new functionality this is sort of moving a little bit further down the Spectrum where it's actually adding functionality not just helping you navigate a big set yeah exactly I think this is part of the core thing right of like the agent interface should match the objective of what you're trying to do right like you don't need a canvas here that is building code for like it doesn't make any sense but like for me when I'm going to a website and websites are like the primary way that people interact with businesses today and honestly for the far future um and so in that sense like if my goal is to understand more about a specific product or a specific company how do I like how does the agent I start to build become a core part of that experience and how does my website move from something static to becoming more of like a dynamic web app so let's push this thought a little bit further should the like should Sherpa even have a website in that like you know it's a bit of a thought experiment should it just be an agent at this point well I think that's like a good question right because you look at e-commerce and part of e-commerce is also like browsing the products that are there yeah and I don't know if I've seen a better browsing experience right it's one thing to say when you are going for a very you have a very specific problem that you're going for and you want to solve that problem quickly right so I think a lot of these transactional uh Services right like a plumber like you just didn't talk agent at that point you don't care about browsing what's there but for something like e-commerce for example like I'm looking for portable battery I'm actually kind of curious at what's out there I Mayan have it like I may be different parts of the funnel right I may know exactly what I want or don't and based on that I need a different interface so let's say for example it's actually serving up like a targeted ad right because I have now given Google my search intent that I'm looking for a portable battery like a certain milliamp or whatever it is right maybe that's the best experience to drop someone right into a product page or even to an agent that's like hey here are like six batteries that we offer that meet your use case because of XYZ um you know click into one of them and we can talk more about it well the other thing too is um this website's beautiful right it shows off their brand uh they've got some beautiful sort of product imagery I don't think we've seen chat or sort of like agent interfaces crack the ability to uh display your brand right on the marketing side of things um because yes like they're very functional but like if every brand just had a chat you know if you went to the website and it was just a single chat agent they're all going to look the same you're not going to be able to get that across and this is where I think like it's getting interesting so I kind of went more down this like web flow Rabbit Hole cool um because uh so the the person who built this his name is Alex he runs a great agency we'll put it in the chat um and he built this on Shopify um and so what I saw there was a couple folks so um like uh uh someone at um here so digital one sec uh his name is Zach's in the vo Community as well like he builds um a bunch of interfaces with voice flow and web flow where he actually uses like the web flow components um to create more of that like agent experience and so we worked with them specifically on redoing our uh like the voice Lo demo form right cuz traditionally the demo form looks like it was like a hubot one so you go in there you put in your information you submit and so we work with Zach say cool can we build an agent that replaces our demo form where you can actually go and talk to it it'll qualify you it'll like go and enrich your information with clear bit um it'll go create all the stuff in HubSpot um but the experience on the user end is actually talking to an Asian to like better describe their use cases ask questions um almost like talking to like a rep yeah and so what he what he was able to do is actually build this right in web flow with our website using all of our core like components yeah so if feels like a really native part of the website here give a give it a little test yeah yeah so like one thing that's interesting that we found with this form is that um you know our demo form um was meant for people booking demos a lot of people were coming to it and the actual completion rate of that form was like 6 7% right it's kind of typical way you see to form so we replaced it with this agent and that completion rate actually jumped to 21% wow um of people who were landing on this page but the other interesting thing was that a lot of people who were coming here were actually just looking for help to like talk to someone now here's here's a question for you do we think think that the 3x in conversion is because this is it's very easy to fill out a form and this is actually weeding out like did we see the top line of uh actual conversions come down you see what I'm saying like a thousand people come in it's so easy to fill it a form that you have uh you know seven only 7% of people go ahead and actually do it because you're like ah you know what whatever versus when you have a little bit more effort in people are actually going to go ahead and complete that form like it essentially gets rid of the people who you know just can kicking right yeah I think it's actually kind of interesting because what we found when we released this was like a the completion was a lot higher but a lot of people actually came were coming to our demo page for support oh interesting so this actually weeded out a lot of the folks who yeah exactly right so if someone comes here and they're like really quickly WR like what is your pricing um if it's going to be able to answer that like based on the information that we have uh in the knowledge base so it's like cool here's our like three plans and then it's like okay cool like I think I need under prise like let me go book a demo um and start actually going and so this is we see where people actually come in they might ask a couple other questions first and then they'll go into being like okay I actually need a demo or I'm actually an agency I want to go to the agency program and so by by kind of like taking somewhere where we know that people are going to and building more of an experience around it we're able to not only like increase the conversion rate but also help people get to where they need to go so the the demos that we are booking um they're actually high quality if someone comes in here and they say like I need support is it going to kick them out to our support bot then yeah good question right now it just kind of says cool like email support voice but we're is we're building in an entire support flow so someone can come in here and ask all the questions that they want kind of like the pricing question but get way more in depth and it'll actually just handle them in this form itself yeah that's interesting that the conversion rate went up um despite because you you would actually imagine the conversion rate if we're getting a lot of like false positives on uh that simple form a lot of them were support you would actually expect the conversion rate of like people actually successfully completing the form to be much higher on the original form yeah right cuz it because it would still be taking in support requests and you know pricing requests all this kind of stuff yeah but the fact that it actually went up as well as triaging folks out means like the actual raw conversion rate is probably even higher than just 3x yeah yeah and like at least for this one we're looking at like conversion rate in terms of like uh completing the form experience right so in this case it would be hitting one of our end pads yeah so they have either successfully booked demo they've been disqualified from a demo um or they've gone down our support path and reach the end of it or they've gone down a partnership path and reach the end of it right so we're looking at like how many people came to this this page started a conversation like with the agent and then actually finished the conversation with the agent I can imagine where this goes and I'm sure your team's already talked about it but like someone fills in a form and immediately like they get an outbound call like from a phone bot hey I'm here from voice flow just want to you know further do some qualification just to really really save our human time yeah yeah 100% And this the goal actually with this was like cool we want to do a couple things right one of them was like I'm like okay I want to replace all the forms of websites with agents cuz we want to be able to get users to Value faster even with that form experience um and the experience today is like pretty broken right where you're filling out this form then you got to wait for an email then you got to wait for someone to reach out it's like let's just get them help now um and so one of the things we wanted to do was turn this into more of like a like a AI like um like accounting experience or AI exp experience right where like you're coming here you're you're talking to a voice Ag and you're able to understand like okay for your core basic questions does this meet your needs yeah um and then if you're qualified being able to do that I will collect more information from you answer more of your questions so by the time that you're actually talking to you've answered a lot of your basic questions and then you and the rep can get into like actual conversation right saves saves a customer time saves the accting exact time and we found a lot of people ask a couple questions on this first before they actually go on oh that's interesting and because it's hooked up with the knowledge base it's unlike a standard form they're able to almost like self disqualify hey actually you know what I I don't even need this yeah 100% um and too like you know right now you can book a demo immediately through here once you get qualified um and we're doing enrichment on the email right so it's like we want to check if an emails over certain point actually if the email we have like three categories so there's obviously the big companies which you book demo right away um then there's these middle ones they're smaller companies but if they're doing a high volume use case they should actually talk to a rep yeah so there's actually a couple questions in here that ask like okay you know um what kind of experience are you building like how many people do you think will interact with it and then we've got logic that says based on the company size if their interactions are high or if they're building have experience put them through the PO rep if not then we're going to put them through a different path they say hey maybe this isn't the best tool for you or go ahead and try our like free and pro PL you can start there uh then you know ramp up and then talk when you review the transcripts are people getting frustrated that they can't talk to a human right away no most people are just asking questions really yeah the frustration comes from more like you know we're making we like release this pretty recently so we're making more improvements to it as we go but um well sometimes people will come right off the bat with a really complex question it doesn't answer in the best way but that's really where we're trying to get better to say okay someone's asking a really complex question like how do we make sure that we can answer it in a way that is like accurate um or perhaps like Trio has a question a bit more to get like a bit more gr so I know we have our Cho support agent that runs like on Discord it's you know runs app as well is this tie is this an interface for that um is this a multi interface agent or is this a one interface one uh agent agent um right now it's one interface One agent so like to kind of show what what BR you're were talking about so why why did he choose to build it with that kind of interface uh just to start off but yeah if we go to let's say like the help channel here um you can see that let's see here yeah whenever someone asks the channel like a question help Channel um our like voice agent will actually go read the question and then create a thread and actually respond to the thread right like in in certain cases actually attaching code um whenever it needs to and so this is able to help get a lot of users like an answer to their question really fast and then you know we'll follow up here make sure things are accurate there's like a rating system in here yeah um but we built this really specifically for Discord um because it's also able to pull things like their role in Discord are they a developer are they not a developer let's let's double click on the rating system does that actually like how does that human in the loop feedback iterate on the actual answer quality oh yeah NCO build on our side we basically like if people rate it well um then that like there's like more weight given to like this actual piece of information for this question cool as it pulls from the knowledge base or is that iterating on the prompt is it iterating on the knowledge or the iterating on the knowledge and like which knowledge to pull for like better questions but he's built a full custom system on the back of this for it and so it's a bit more involved but at the beginning it was like human the loop right where we go in and be like okay these answers are bad like just because information's at a date we need right when we release voice we have to release we update a bunch of stuff so we've got a Discord agent we've got an inapp help agent we've got a lead gen agent these are all running off of you know sort of single agents with a single interface do you see a future where you have a single totico agent with multiple interfaces yeah yeah why have we not taken that approach today um well this one we just wanted to get something out quickly yeah but realistically like the knowledge base we probably need for this will be the same knowledge base that we have for the Discord bot because it's been like really refined over time yeah and so yeah like that knowledge base should actually be shared between like our in and this and Discord um so the actual building so the knowledge makes sense but the actual building of the agent though yeah do you foresee do you foresee most folks like let's say like two three years out being able to build a single like agentic prompt that is able to handle multiple interfaces or do you think that folks are going to really segment agents and interfaces for you know a while to come um I think at least for the short term like just being honest like we found it easier to be able to build specific agents for specific interfaces because sometimes you like this one webflow has a bunch of custom actions that are tied to web flow components yeah right so like this thing over here like hi I'm too voice flows agent um is actually and this one up here are actually custom actions and voice flow that are like sending basically information to the front end to render different UI elements that's cool so if I like go through here right and I go through this like book of demo flow and I go through the thing I did last time so Daniel and Daniel boy.com um you'll notice that like it's going through a bunch of enrichment right now in the back end but it's going to then pull up my name and then pull it into this that's probably where that latency is coming from then cuz you're actually setting that the clear bit you're pulling company size any rule information yeah yeah we' actually we're sending clear we're riching it we're actually also creating a containg HubSpot oh um and so like now that we've added more stuff there's probably more like thinking stuff we can probably add behind the scenes here yeah HubSpot for folks who might not know it's just a serum we're using on the marketing side yeah but you can see here like thanks for providing your name Daniel let's get started um because it pulled my name into this custom action and then on the top here it actually moved from like info collection to or you know identifying to info collection it's a nice I always love uh uxs that show the customer sort of like their progress yeah because you don't want to go through like an endless number of forums but like okay I can go through a little bit of pain here because I see I'm on step two or three I'm almost done yeah 100% And if I go through this to like okay like I I'll just be really quick with the customer support agent and then it's going to ask me like how many um like interactions or how many people interacting with it great use case Daniel how many people uh let's say like probably 10,000 people um it should go through uh actually maybe it needs to be higher let's see um Okay cool so here disqualify me um but uh if I had maybe higher interactions or my company size was bigger um it it would actually it actually pulls out a component that says cool here's like a quick summary of what we talked about um this is your email is correct and it's an edible component so you can actually go in and change the email if you want yeah um but like for this interface there's a lot of custom actions that are tied to it because we want custom UI elements in the front end to be appear right for Discord it's very different because instead we're doing a lot of just stuff the the Discord API and so for right now like it kind of makes sense to have an agent for interface because you want to be really tightly coupled to that interface yeah but I think as time goes on it does make sense to make it more um like one agent multi interface we just have to figure out like you know different front end things happen in different faes so how do we make sure it happens on one side and not the other yeah I you know my my question was a bit of a leading question as um yeah I often you know I I think in our industry you see a lot of companies where they talk about like you build one agent one you need to deploy everywhere right now the challenge is like do you think about each interface you know interface kind of being like synonymous with channel here like WhatsApp SMS email Etc it's almost like a vend diagram where each channel has its own native capabilities right so like apple business chat has native capabilities what'sapp native capabilities voice etc etc the only thing like if you kind of listed all these Vine diagrams the only common denominator is text right and so when I've uh you know companies come in and say Hey you know sell and so company talked about build one agent deploy it everywhere yeah they're either talking about a really simple agent that is just text or they're talking about a Frankenstein build experience behind the scenes where you have to add in all the little exceptions per Channel yeah and so we've actually done work like this before where uh voice and chat can be done a little bit more easily but you actually have to set like uh if you want to use all the capabilities of chat for example carousels uh buttons things like that you actually have to change up the LM prompt per channel so that when it comes in you have a additional prompt you say hey look at the interface if it is uh this channel use this prompt if it's this channel use this prompt cu the other thing even voice like you might want to use bullet points so yes it's still chat but like on chat you want to use bullet points on voice you don't want to use any bullet points right you want to be a lot more conversational so you end up having a completely different experience and I think in a few years as models get better you might be able to start prompting in a way where it can actually handle all channels in a single interface cuz you're having to respond hey if this is voice respond but this sort of format if this a chat respond in this sort of format but if you're doing more than two channels right now it would just blow up the prompt and you're just adding so much unnecessary computation into every single step versus just building things separately and having some reasonable components right like maybe it's some sort of API call that's reasonable hey I'm just going to create a component for this but I'm going to go build like my prompts for chat my prompts for voice and they can both pull in that like off component or something like that yeah I think that makes sense um at least for right now too I think even for us right like as we work with like people who are in on these interfaces like a lot of people are still trying to figure out like what is the best way to do this and so for right now it does make sense just to have like one agent focus on interface and make that really really good before we do anything else and like to give you a sense of what's Happening behind the scenes here like let me just move some of these closer together um like there's there's a ton of stuff happening like in that demo form from like capturing an email we're upserting Hub spell contact we're then like actually enriching it with clear bit we're running logic to figure out like which path to we send people down and then here's these custom actions that I was talking about right like we're basically sending like a progress update to to the front end to then change the stage based on where someone's at um we're like you know basically writing a custom acknowledgement message that's then sending a payload to the front end as well so there's a lot going on that's very tightly coupled to that front a lot of this is interface logic it's not conversation logic yeah a lot of this is interface logic like the actual conversation itself is like fairly straightforward right we like right now we're just asking a bunch of questions but in other areas it's just a loop with AI where it's like answering their question um and so a lot of this is actually built around the interface because we want to basically keep people in a loop to answer their questions and then as they go down a certain path like booking a demo we're sending them more on a straight line where we want to connect specific pieces of information and based on certain points in that conversation we're using that information to trigger some sort of front end component or front end UI yeah right and so actually like what's interesting too is like I think more and more will happen is like like my thought is that a lot of these like a lot of agent Builders will actually move towards being like um interface orchestrators yeah right where the actual underlying model is handling most of the conversation um a lot of the API calls and figuring out what to do and where you actually need to integrate with a front end is is that's where you need an orchestrator right to be like at this specific point or if users said a specific thing we want to render these specific front end components and here's the information we want to be able to pass to in front and like you know that can be really basic but I think as you get more complex and you want to have a lot more control over that like frontend UI experience so especially when it's something like a website and you're interacting with a lot of different components in the front end yeah that's where like an orchestration type of platform for that it's like very valuable so do you think the work should be done on an agent orchestration platform like voice flow or do you think that website Builders are going to start introducing some of this orchestration yeah good question I think it's both I think they'll both kind of go towards it right like because web is already like adding more AI stuff in um to their website but I think when it comes to this this like interlayer between like your backend Services um and your front end which in this case would be a website like I do think that there's going to be something in the middle because like the website Builders themselves can probably handle a lot of simple basic use cases but once you got into like the more complex territory of like I have custom back service I want to integrate with like I want to be able to handle a lot of different business logic like that's where you actually need like an orchestrator right and the question is are you building this in code yourself or using something that's like easier for collaboration like a builder yeah and I think that's always going to be the conversation you're going to ask yourself especially when you move towards like more complex production agents um is you know do I rather this encode myself or do I use some sort of build so I can make it easier to collaborate with people on it it's going to get really messy between what is an agent and what do we just consider like a core app today so good example of this is like um let's say uh you know website uh that's build webflow you know over the past 10 years you've seen a lot of like AB optimization tools where like it might change the copy based off where you are like simple little things right but what if the actual website itself is starting to change as you use the website right it profiles you know okay hey this is a hobbyist user or this is an Enterprise user the actual website itself is going to change dynamically at what point is that not just almost considered an agent right behind the Stream if behind the scenes a lot of that like orchestration of the front end is being done with LMS it's almost an agent in that degree right like it gets a little messy between what is what is an agent and what is an app at some point they're almost all the same thing but it might look like a website well that's the thing right and I think like that's where my initial point was around like I think websites are changing to web apps yeah um and I think honestly apps are going to be synonymous with agents because agents are just a core functionality of an application right like it's just changing the way that you interact with it changing the way it's interacting with the backend system um and making the experience over a lot smoother a lot more personalized you know an example I've given before is like I would love if Netflix was almost a little bit more like an agent um we've all had the experience where you're scroll you know you're scrolling trying to find something you like there's too many options by the time you get to the bottom of the page and you go back to up to the top it's almost a conversation that you know it's implicit conversation that you have sent to Netflix saying hey I don't know what to watch yeah right at that point refresh the UI suggest something you know when you get back up to the top of top of the page it's a completely new set of items like that's a conversation to some degree I'm just not saying anything or even you know with Shopify there's a lot of times where uh maybe you're searching and you just can't find the right item whatever it is that is a implicit conversation that you're having with that app yeah where that app if it was a little bit more agentic says hey look at their profile look at the items that they've looked for they keep scrolling around clearly they can't find what they're looking for or they don't know how to do something yeah yeah I think it's a really good point it's not like conversations can be implicit or explicit I that's actually pretty interesting um but yeah like I think like in terms of like an agent versus an app like at a certain point it's Al mantics right because it's like at the end of the day you're build the user is trying to accomplish something um in Netflix it's like entertainment right I'm trying to find a good way to spend time I want to be able to watch a show and your app is just like your interface of like what is the most effective way to get you to that end goal yeah um and like an asent is just the orchestration of how this app is working yeah and so you know regardless of what words we use like at the end of the day these are all like applications and it's figuring out like for the interface for the UI experience like what the user is going through like what is the one to be able to help them get to their objective yeah um and so yeah whether it's Netflix and it's a visual interface whether it's completely voice whether it's completely chat um I think that's something that we're going to figure out but that's very use case specific yeah so I want to kind of wrap up on one topic is we're kind of leading into it which is like co-pilots as an interface because I think you're starting to see and when I talk about co-pilots you know I think what you showed with Sherpa is to some degree a co-pilot right he's trying to help you do something yeah but I think what um is really interesting is a lot of the co-pilots that are launching where it's actually helping you do a very complex action and it's actually taking some of the work off of you the user right so a lot of the stuff we're showing here is information yeah but it's not maybe like you know doing something um repits a great example where it's a essentially a an IDE that's browser based what's really cool though is they have a co-pilot where they've built this awesome IDE over the past I don't know five six seven years now but now they have a co-pilot experience where you're actually using an agent to do the work on top of this platform for you right and it really changes what you could even think of as an agent from like informational to this is the way to actually get work done on top of the platform I'm using you're seeing Salesforce do this repats now doing this cursors like the entire experience is really a co-pilot it really changes how you think about sort of agents in general from informational to this is actually the core way that I interact with the app itself yeah that's what you're referring to right like this cobal on the side like there's this functionality and then obviously they've got their like agent Builder as well or like sorry um what's it called their rep agent right yeah they they have two they have like rep agent and they've repet assistant right and assistant is helping me understand the code and like agent is actually building the code for me yeah it completely changes the way you think about SAS apps the more you use these types of tools yeah where you know I know today's sort of topic is on interfaces but diving a little bit more into these like co-pilot style interfaces it completely changes what the app is so it's kind of like SAS 1.0 was like you have a problem and you have uh you know like a software that solves a very particular problem right like a point solution SAS 2.0 sort of like a platform right we have a set of tools like a browser based IDE allows you to solve many many different problems but you have to do all the work these co-pilots are going a step further in kind of this like SAS 3.0 where yes I have a platform that gives you observability and the ability to actually go make the changes yourself do the work yourself but the AI can also just do the work for you right like you don't even have to like the platform's still there if you need to make changes but as you can see here it's able to go ahead and generate code it's able to actually go build out a SAS app and then you can always go into that like SAS 2.0 platform interface to go make changes yourself understand how it works yeah but a lot of the he heavy lifting is actually done by the co-pilot itself which is totally different than a lot of the Bots you see today which are purely just like information exchange yeah 100% yeah I think this is super interesting like like repent agent Builder is like fascinating because like I I use this and I use cursor uh quite a bit but being able to actually yeah you're right like show you how to use the tool itself and then actually do it for you um is is super interesting and and I think it goes again to the point of like what is this application what is it trying to do like I'm using replic cuz I'm I'm I'm I'm building in an app and Facebook whatever it is right and so it's a builder and so it makes sense for this agent to actually build it for me yeah and then the question is like what is the best way to show you how to do that while still giving you that that aspect of control and understanding what's happening yeah which is why I'm using a tool like replic or something else but you've seen there's you know LMS are really good at coding there's just there's so much data on uh uh so much sort of code data out there for them to ingest and it's also really just structured logic which is where often these LMS really Excel so I think you're seeing a lot of way is going to transform the rest of this ass industry take place first in coding tools right cursor co-pilot uh Reet agent um Devon's another interesting one and you see a spectrum on these tools where um you have more of a cursor style where it is there to assist you and it can do some of the work for you but you still that full full observability it's still very much IDE all the way to the other end of the spectrum like what Devon uh is trying to do where it just does the whole thing for you like there is less observability the agent like you give it a task and it just goes and does it right and you're seeing this spectrum and it's going to be interesting because I think Devon is probably right 5 10 years from now but I think when the LM models aren't as good uh people are sort of preferring to cursor approach which is why that company's just completely just blown up recently right they've done incredibly well U and I think it's people still want to see what's happening behind the scenes they want to work in increments yeah versus that whole just do the whole thing at once for me yeah yeah as good point yeah I think at the end of the day it's like right interface for the for the for the goal yeah right like I I use cursor over repl agent because um I want that level of control well also the right interface for the current moment in technology right yeah that's true that's a good point there's um you know if you go back to that Spectrum if AI could do exactly what you needed to do perfectly every single time Devon is the right approach right just put as minimal overhead on the user as possible but until it's able to do that you need to sort of have this incremental approach where you have all the visibility in the IDE and you're able to increment with the AI sort of hand inand but I think you know just again looking at where models two years ago were to where they are today I think you can see like Devon's got the right approach long term but the co-pilot's working at increment right now cursor repet I think is is the right right interface for sort of the moment in technology yeah yeah yeah it's a really interesting point um well I think next week we're going to be chatting about the death of SAS quote unquote uh with agents and things like that we can probably end it here talk about interfaces uh but hey hopefully this was interesting for everyone we want to do sort of a casual riff based approach just talking about a different topic every week right uh so this week was interfaces uh next week where we want to talk a little bit more about how SAS is changing we start to get into that a little bit at the end today and uh we're going to be doing this every week humans talking about agents so thanks for tuning in thanks Daniel for taking the time and we'll see you next week everyone [Music]