AI Dev 26 x SF | Atai Barkai: Fullstack Agents & Generative UI with AG UI

Added: 2026-05-23

[00:00:07]Cop of copet kit. We make developer infrastructure for uh copets which are user interactive agents as opposed to fully autonomous agents. And we really make the layers that connect the agentic world with the userfacing application world. And today copc powers tens of millions of interactions in production uh every week between agents and users from companies like Cisco and docyign and function health and dosia telecom.

[00:00:31]And in fact the majority of the fortune 500s today uh use copag and ay uh driving you know millions of weekly installs. Uh we're also the company behind the agy protocol. If any of you guys have heard of it, it stands for the agent user interaction protocol. Uh I'll talk about it a lot more uh soon but this is a protocol that was born from our partnership with lang chain and crewi uh and since we launched in production it's been adopted by almost all the top uh agent backend agent harness builders including Google Microsoft Amazon Oracle and many of the top AI startups including Mosrap Idantic Agno A2 Landex uh and others um so today I'm really excited to talk about the state of Aentic UI where that whole world is going uh to give a quick technical overview of this world and then talk about the generative UI spectrum of the different solutions in this category uh and finally talk about where where are things heading in this category.

[00:01:29]So first the reason I'm really excited to be building here uh and to be presenting about this is that we believe over the next years all UI will be AI uh which means that you know over the next years all interactions between humans and technology going to become increasingly mediated by hygienic systems and of course that's true of the complex software we all use like uh Figma and HubSpot and of course cursor and cloud code uh but it may even be true of your refrigerator and The first time I heard that from one of the large electronics manufacturers that approach us, I thought it was a little bit funny.

[00:02:04]But on second thought, wouldn't it be nice if you could walk by your refrigerator uh and just tell it to order the ingredients that are missing for lasagna tonight? Uh so that's the world we're heading into. All interaction between humans and technology mediated by agents. Uh but what is it going to look like? Well, I can we can we know what it's not going to look like, right? It's not going to be giant boxes of black and white text that we're all going to be staring at all day long. Uh you can draw really interesting parallels between the initial state of AI UI and the MS DOS interfaces of the you know mainframe era of computers. Uh and you know these initial crude interfaces are great for early adopters like many folks in this room to extract enormous value out of these systems but mass adoption uh really requires AI to have its Mac and Windows moment. And that's exactly what's began to happen in 2026. We're now in the transition from the MS DOSS era of AI to the Windows and Mac era of AI.

[00:03:06]Uh and so that's where we're going. Uh but as many folks in this room probably know building great full stack agenic applications is very very challenging.

[00:03:17]Uh and the core reason if you dig there is that agents break the request response paradigm that's been powering the internet since its inception for the last 30 years. Right? So agents are of course just uh software right? But from the standpoint of pre-agenic software they're a little bit like a weird bird right? They have to uh they're long running so have to stream their work as they're running. They have to support reconnections uh and uh even midrun steerings. They also need to support structured and unstructured data exchange at the same time. Right? So you have tool calls and state updates at the same time as uh text and voice. So you're almost like you're building a traditional application and Slack and Zoom at the same time. And there's really a long long list of these peculiarities from the standpoint of pre uh agenic software which make it very challenging for builders to build great full stack agents. Uh and that's exactly where copagit and agy come in. Right? So copit provides open-source developer SDKs for building uh these full stack agents and a cloud self-hostable cloud that is optimized for the agenic paradigm as opposed to retrofitted from the request response paradigm. uh and AGI the agent user interaction protocol uh lets developers who build aenic backends and harnesses uh like a companies I mentioned before bring their agents into user fake facing ecosystems by building against a robust standard right so they just conform to a few simple requirements and everything just works and continues to improve as the ecosystem improves uh so uh you know agy our initial focus was web frontends But there's a quickly growing uh list of AGI frontends for other stacks including mo mobile, WhatsApp uh uh text messaging, even voice.

[00:05:07]This is a diagram of uh the architecture of Copek and AGUI. I actually don't have time to dig into this uh uh in detail right now, but you're welcome to take a picture. I see a lot of people doing this and if I have time at the end, I can ask question I can answer questions about it. You can also stop by our booth right over there. I think it's 500.

[00:05:27]Uh, all right. So, now on to generative UI. First, what is generative UI? Uh, generative UI is the set of UI paradigms that is both enabled by LLMs and aenic systems and also is a solution to many of the problems, new problems that agentic uh systems introduce to software. And after working with the leads here for the last two years, it's become very clear that uh there's no single generative UI solution that is appropriate for all contexts. And really generative UI solutions exist on a spectrum that goes from fully controlled on the left side uh to fully open-ended on the right side. And really every single point alongside the spectrum is appropriate for a different context. And in the remaining of the talk, I'm going to very quickly go through the three major pillars of the generative UI spectrum, which are controlled, declarative, and open-ended generative UI.

[00:06:25]Uh so first, control generative UI. In controlled generative UI, developers define uh a collection of pre-built, predefined components that the agent can then call upon and use uh on demand. So now I'm going to try to do a live demo.

[00:06:41]internet has been a little bit spotty.

[00:06:42]So, let's uh pray for the demo gods.

[00:06:46]So, there we have a user asking for a breakdown of revenue by category, right?

[00:06:51]So, instead of getting this uh impenetrable long paragraph of text, you get this beautiful pie chart instead. Uh and there's the same idea if you wanted to see it in bar chart for example.

[00:07:03]And these are all fully interactive components that developers define. So, let's look at the code behind control genui.

[00:07:11]Uh this is using the copcet react front end. Uh so we define a hook called use component. Uh we pass it the name of this component when it should be invoked by the agent. In this case when the agent wants to display a pie chart the parameters that the agent requires to get from the agent to display this component. In this case the title the description and the data inside of this pie chart. And finally a react component to actually render it. Uh here we're using by the way uh zod which is the typescript equivalent of uh pyantic if you guys are uh not familiar with it.

[00:07:45]So control jiu is in many ways the most boring part of the generative UI spectrum. Uh but it gives you pixel perfect control of the designs. Uh and it's the most deterministic of a whole bunch and that makes it a really great fit for the few most important surfaces in your application. Uh so for example, if you're an airliner, you want your flight tickets to always display exactly in the same way and to be as deterministic as possible. Uh so for that reason, we think of controlled generative UI as the workhorse of the generative UI universe.

[00:08:21]Uh next on the spectrum, we have declarative generative UI. uh in declare the genui developers uh declare a catalog of building blocks Legoike building blocks that the agent can then assemble on demand to answer and use a query. So let's see that on demand as well.

[00:08:41]So here we have a user asking for a sales dashboard based on some data that the agent can query. Uh and you can see the agent is assembling this sales dashboard on demand, right? So every one of these subcomponents was predefined by the developer ahead of time. But the but putting them all together in this form was done by the agent on demand. It was never defined in advance.

[00:09:04]Uh and there are a bunch of different solutions on the market today for declarative generative UI. Uh of note there's A2UI from Google which I'm going to talk about right now. Uh there's uh JSON render, there's open UI, there's hash brown. Um and they're they're very similar to one another. Uh A2I is again a protocol uh spec that was led by Google with major uh involvement and code design from Copedicit. We actually co-launched with them version 0.0 of this uh less than a week ago with major major improvements. So let's look over the code here. Uh so the first thing you do is you declare the catalog definitions of your components. Right?

[00:09:43]So each one of the building blocks, the legolgo like building blocks in your catalog needs to have a description when it should be used by the agent. What data is it expecting to get injected from the outside? And you repeat this for every single building block in your catalog. And then the second part of the catalog is the renderers, right? Where every one of those building blocks now gets a render in the native environment where your front end lives. In this case, React once again. Uh and essentially that's it, right? the collection of catalog plus renders defined that catalog uh uh entirely and whenever you use an any AGI compatable agent which again is almost all of them uh you can you can teach your agent to speak an AUI by simply passing it as um as a parameter in your catalog.

[00:10:31]Let me try to just put my headphones again and see if can hear myself here.

[00:10:36]Okay, that is actually a little bit better. So going to do a DJ session.

[00:10:42]Um all right. So with declarative generative UI, you're giving up a little bit of determinism because the agent is the one assembling components and a little bit of fine tuning of your UI because uh every UI that is rendered needs to be an assembly of pre-built components. But in return, you now get the agent to be able to generate uh uh solutions to any user query that you can think of. And that profile of trade-offs makes it a really great fit for the long tail of userfacing surfaces in consumer applications uh and also for internal enterprise applications, right?

[00:11:21]Where the efficiency of the implementation is really important and now you're just defining the catalog once and that's it.

[00:11:28]Uh so finally at the far end of the spectrum we have open-ended generative UI and there are two uh variants of note in this world of open-ended genative UI.

[00:11:37]The first is fully open-ended generative UI right where the agent is actually writing JavaScript or HTML code inside of a secure sandbox on demand to answer a query. uh in this case we have for example uh asking the agent to create a new brutalism themed calculator right something that I don't know if ever existed before and it's actually building it on demand and it's usable for us to to use so that's one variant of open UI the second variant is MCP apps um and this is a standard that was adopted by um openai and claude into their own app stores um where applications can expose expose many applets to be embedded inside the chat experience. So here we're going to try to uh again pray for the demo gods to see the excaladraw mcp app embedded into our chat.

[00:12:30]And as you can see this is a little bit slower. This world is a little bit still getting started but we asked here to see uh a network diagram of a router connected to two switches and two computers. And you see that actually built that in a real Excal right we can interact with this go to Excal and so on. Uh from a code standpoint, it's extremely easy again to enable both MCP apps and fully open Gener UI in any AGI compatible agent. You just pass the URLs of the MCP apps you want to support. If you want to enable open-ended genui, you can do that. Uh and that's it. Uh so MCP apps are really great for third-party tools, right? It's optimized for the super hosts of the world for uh uh chach for claude maybe for cursor. Uh but with the AGUI MCP apps handshake, you can bring the same MCP apps that your users are familiar with in from the app stores to your own uh custom full stackenic applications.

[00:13:25]Full open-endive UI uh is really great for fully custom visualizations that uh you can never see before.

[00:13:34]So as I mentioned, every one of these solutions has its place in the stack, right? So to quickly recap, controlJV UI is where developers define a collection of pre-built components. It's pixel perfect. It's mostly it's the most deterministic of the whole bunch and it's great for the most used surface of your application. Again, if you're an airline, your flight tickets components.

[00:13:54]Declare genui is where you declare a catalog of building blocks and the agent is doing the assembly and this is really great for the long tail of user facing surfaces and for internal enterprise applications. Again, fully open agenda UI is great for third party applications. you want to embed in your own application uh and for fully custom experiences that nobody could have foreseen ahead of time.

[00:14:17]So the GenVI spectrum is really just one part of the full stack agent experience and I want to take a quick opportunity to walk through one part part that ventures out of this general UI paradigm uh and that's shared state. So first let's see an example of that where the agent and the user can collaborate over a single piece of state. So, of course, we have the the classic hello world, the to-do app. Uh, and the user can ask that agent to create a list of to-dos. The user can then manage these to-dos and ask the agent uh to operate over whatever is remaining. So, for example, you know, what should I do for the remaining open to-do and the agent can see exactly what's going on.

[00:14:57]So this is a really great abstraction if you're trying to build your own cursor for any type of uh use case.

[00:15:05]From a code standpoint, it's extremely straightforward. If you use again any AGUI compatible agent, you use you get access to the use agent hook. And this use agent hook gives you access to agent state and agent set state. So just like you have a standard React local state in any traditional application, you can use exactly the same abstractions with a state that is synced both directions with the agents with all the different performance uh uh issues and uh synchronization issues automatically taken care of by the stack.

[00:15:39]So where's the world of agent user interaction going? Again, we like to think of Copic as a company behind the agent user interaction step. The first part of the stack is enablement, right?

[00:15:49]The technology that allows these agenic systems to be built at all. And of course, that's where the majority of the focus of the industry is right now as these applications are still getting built. But once you actually build these application, deploying to production, you you ask yourself what are your users doing with your genic application in production, right? Because as soon as you've shipped your uh agent, you've broken all of your pre-existing user analytics. So you have the insight layer. Uh and finally what we're extremely excited about is the final layer in the stack which is continuous learning from human feedback. Right? So again today everybody's building the first agenic applications. Many folks are helping uh are choosing us to help them do it which is very exciting. But the next thing that everybody wants to do is get the agent to be smarter right more autonomous make fewer mistakes u be higher quality and that's really really hard to do. And essentially all of the approaches that work in this world require very high quality data of you know what did the agent get right what did they get wrong uh today many companies get this data from human labelers right and I always like to ask who's ever labeled data for company like scalei or americanore so almost anybody no almost nobody there's two hands up here who's ever used cursor so almost everybody right so cursor is actually using all of your interactions with the application to train frontier models for coding, right? Every time you accept or reject a suggestion, every time you edit something the agent did, that's extremely useful signal for what the agent should have been doing instead. Uh, and we're really excited to bring that technology to any AGUI compatible agent. So, anybody building aic applications can have the agents learn from the experts who use those applications every single day.