Full transcript
0:13Hey, everybody, thank you very much for joining us.
0:18My name is Alan.
0:18I'm joined by several other peers.
0:20This is a partnership conversation
0:23with a lot of different players talking about generative UI.
0:26I'm very excited to be here.
0:27We have a lot of good content.
0:28We're going to try and make it through all of these slides.
0:31We're going to talk about my and many of your frustrations
0:36with a text only interface, and how we believe
0:39that all of the interfaces will soon have some sort of AI
0:43behind them, and you can do a whole lot more with it.
0:45But I'm actually not going to start this talk.
0:48I'm going to hand it over to ido to come here and talk about MCP
0:51apps to get us started.
0:54Thanks, Ben.
0:59So hi everyone.
1:00Thank you.
1:01Alan sorry.
1:06So I'm Ira Solomon.
1:08I am the creator of MCP and agent craft
1:10and co-creator and maintainer of MCP apps.
1:14Before we get started, I just wanted
1:15to shout out to the amazing people that made MCP happen.
1:19So these are co-creators.
1:20We have liad and Olivier and Anton and Nick
1:23and a bunch of other great people.
1:25So as Ellen hinted at, is text isn't enough.
1:31So what is.
1:32How do we transition from text into rich graphical user
1:36interfaces.
1:37That's exactly why MCP apps was created.
1:41It's the official extension to MCP.
1:43It's how we do interactive apps over MCP, standardizing
1:47transmission of UI to clients and the communication protocol
1:51between clients and UI.
1:54Let's see how it works just to get an impression.
1:58So, for example, let's say that I have a product
2:01and I want to understand the status of my funnel.
2:04So I ask my agent and in the old world
2:06it would reach out to my analytic provider,
2:09like post hoc, and we'd get back this textual response, this wall
2:12of text, which is nice and factually correct,
2:15but doesn't really help me understand
2:17what the hell is going on.
2:19Luckily, because both Cloud in this instance and post hoc
2:23support MCP apps.
2:25Instead of that, I can just say, show me.
2:28And now instead of getting that sorry, instead of getting those.
2:39Sorry clickers messed up.
2:47OK, so because they both support it, then we can say show me.
2:50And now instead of text I can get this interactive UI
2:53component created by post hoc with its own branding and user
2:56experience, and I can see what's going
2:58on with my funnel in a glance.
3:02Now, we don't want to be constrained only
3:05to things that are predefined in some application.
3:09We do want to let the agent run free
3:10and create a bespoke UIs for us.
3:13For example, let's say that we want
3:16to understand what a funnel is.
3:19So what we can do is just tell the agent,
3:21explain the funnel to me.
3:23And now instead of writing something
3:25that I need to read and understand,
3:28it could just generate something on the fly.
3:30This nice UI explaining to me what a funnel is
3:32and how it works.
3:34Now, this isn't just presentational.
3:37As we said, apps also supports interactive interactivity.
3:40So actually embedded into this UI component is interactivity.
3:46So let's say that I want to understand more
3:48about a particular phase in the funnel.
3:50I don't need to prompt it anymore.
3:52Don't need to figure out what I need to say to get what I want.
3:56The model already embedded a behavior
3:59that when I click on, show me one of these steps,
4:04it will actually send a prompt to the agent
4:07to drive forward the flow and give me that
4:10explanation that I need.
4:13So we saw technically what MCP apps
4:16is capable of but if we Zoom out for a second,
4:20it's not just nicer chat.
4:22It's actually a fundamentally new way
4:24to interact with the web.
4:26If we look at the old web, for example,
4:29even doing something simple like preparing for a special occasion
4:34would require me to go to 10 different websites
4:37and understand how do I book something, how
4:40do I order some flowers.
4:43How do I do a bunch of stuff.
4:45And if you look at these apps or these websites,
4:50they're really complex.
4:51And the reason that they're so complex
4:54is that they're meant for us humans.
4:57And you don't know as a website, what
4:59do you mean is actually trying to do.
5:01So you just push everything that the human might
5:04do into one screen, and you need to simplify it or build it
5:07in some complex way.
5:09So you'll be able to consume it reasonably.
5:14But if you take these complex UIs and break them apart,
5:17you suddenly realize that agents aren't like humans.
5:20Agents know exactly what you want
5:22to do so they can take just what they need.
5:25Just a map, just a particular checkout button,
5:29whatever, and present it to me directly.
5:32So in this example, all we actually
5:35needed to get this flow going is the map
5:38to see where it is and maybe the hotel
5:43and what that flow would actually
5:44look like in an agent with MCP apps is
5:47that it would surface the date for me,
5:50just the UI of that date.
5:52It would surface just the shopping aspect.
5:55So just the flowers that I want to buy.
5:57Same goes for the hotel itself, but I
6:00would say it's not just limited to predefined applications.
6:03We also want the agent to be able to run wild and create
6:06bespoke UIs just for what I need to really complete that end
6:10to end flow.
6:14So if we look at what we just saw,
6:16there's kind of a spectrum of UI like we have the predefined UI,
6:19we start with post-hoc.
6:21It's just an app.
6:21It's just like a website that you've built until today.
6:26The second step is going towards the decorative UI, which is more
6:30of you describe basically what type of components you want,
6:33and the host builds it the way that it sees fit.
6:37And the third one is what we saw with clouds.
6:39Generative UI is really have the model
6:42create content from nothing.
6:44Just stream that HTML into the Canvas, into the app
6:47and display it.
6:49The beautiful thing here is that MCP apps doesn't care.
6:52MCP apps is agnostic to what's being rendered.
6:56That's actually what makes it so powerful in interoperability
7:00with other protocols.
7:01And as you're going to see here soon,
7:04you have a UI and UI that are practically built into it
7:08because they can be rendered inside the MCP app
7:12and conversely, they can be MCP app can be used as a component
7:16that these frameworks, these protocols can render.
7:19So you can have really rich experiences in the chat.
7:24Now MCP apps are everywhere.
7:25I don't know if you're aware of it,
7:27but basically every major host supports it.
7:29So we have OpenAI and Cloud and Gemini
7:32and a bunch of others just looking
7:34at ChatGPT and their apps.
7:36They have over 500 apps from huge companies
7:39like Walmart and Salesforce and you name it.
7:44So this isn't something that this is just a concept.
7:48It's much more than that.
7:49It's just the new app distribution surface.
7:54If you look at, again, ChatGPT, for example,
7:57they have one billion weekly active users,
8:00which is 12% of the world's population.
8:02If you add Gemini and Claude and other hosts,
8:05the number of users that are using chats is much bigger.
8:09That's over 160 times the number of users
8:12that the Apple App Store had when they launched.
8:16So as you can see, it's a huge total addressable market.
8:19And the question that you need to ask yourself now
8:22isn't why you need to develop an app.
8:25Why do you need to get your product
8:27ready for the gigantic web.
8:29It's what happens when users stop being
8:32the people that see my website.
8:34It suddenly becomes agents.
8:36So I hope you look at this thing as a unique opportunity.
8:40Just to take a step back, look at the products
8:43that we're building, take them apart and build them back.
8:47Not as a monolithic app, but actually
8:49as part of a web of applications connected together by the models
8:54and with MCP, apps and other protocols like E2E and agwe,
8:59you can really build your product once
9:00and run it everywhere.
9:02Thank you.
9:14Thanks ido.
9:19Atoui before I get into that, my name is Alan blunt.
9:23I was a full stack web application
9:25developer for 20 years and then did some machine
9:27learning stuff in Google Cloud, and now I'm
9:30doing agent building platform.
9:33And a lot of the protocols that you have heard of atoui
9:36is one of that list.
9:39We are actually a hybrid team collecting
9:41a lot of people that have contributed
9:43for over two years towards this initiative is
9:47what you are seeing now.
9:49This team is trying to solve basically
9:53these three sets of problems.
9:56When I create an agent on my application
9:59and put it on my front end, it's a little bit easier.
10:01But if I have a remote agent that is part of my system
10:04or a mesh of agents, how can they communicate.
10:07How do I do true generative UI.
10:10And how do I get that to users safely, wherever they are.
10:15It's sort of a complicated action space,
10:17but this is what we're trying to solve for with atoui.
10:21Let's start with remote agents, and let's start with a diagram.
10:25This is just one configuration.
10:26There are many configurations, but it sort of sets the stage.
10:29So I've got my design system, my front end application
10:32which is where my users are clicking and doing and typing.
10:36That front end is connected to some backend,
10:39some backend server.
10:40That is my APIs, my agents, my whatever.
10:44That's a very normal pattern that we all have.
10:47But maybe I have these agents that
10:50are not in my system, that are in some remote system.
10:52That is a particularly hard part of this equation, especially
10:56when you get into users and ACLs and all
10:58of the other parts and pieces of the system that
11:00need to be secure.
11:01So we really needed a way to do message passing
11:04across these entities in a way that we could secure to treat UI
11:09as if it was just data.
11:12So we don't need all of that.
11:14Let's start with a simpler scenario.
11:17If you have a front end and you have a back end,
11:20and maybe that back end has some way
11:21of generating a UI, an UI message,
11:26you could just ship that across.
11:27It doesn't matter if you have subagents
11:29that are local subagents you're in the happy simple scenario
11:32path totally works.
11:36Ooh, I did want to point out that transport
11:38is totally agnostic.
11:39You don't need a two way you can do it over rest.
11:41You can do it over web sockets.
11:43You can use aighewi.
11:44You can use MCP for your front end.
11:46It doesn't matter.
11:48But sometimes you may collaborate
11:50with other remote agents.
11:52Those remote agents are the ones actually generating
11:55that E2E message, and you're just
11:56passing it across through the orchestrator because it's
12:00just a normal message payload.
12:02And we'll get into that in a second.
12:03It's very easy to simply pass across.
12:06One other permutation is the remote agent
12:09may be doesn't know how to generate an E2E message,
12:12but your orchestrator does.
12:14You can manipulate or mutate that message
12:16as it comes across if you need to.
12:19These are just a few of the patterns.
12:20There's many, many, many others, but I wanted to set the context.
12:25Now the generative UI part, this is not using tool calling.
12:29This is just structured output.
12:31We chose json-ld because many LLMs,
12:34most LLMs are pretty good at JSON.
12:37There are other types of declarative UI formats
12:39that we could use.
12:41We picked a thing that worked for us today.
12:43It could certainly evolve if we needed to.
12:47The LLM assembles these components together.
12:51We'll talk more about the components in a second.
12:53These symbols, these components together,
12:55maybe one component, maybe multiple components
12:57into a widget or a surface is what we call it,
13:00some kind of a layout.
13:02And then it hydrates those variables in that surface
13:06or populates the variables.
13:08These are streaming chunks of messages.
13:12So it's a pretty good user experience.
13:14And again, very flexible.
13:16And one of the use cases I'm very excited about is cacheable
13:22configurable UIs. So your UI is already configured full UI
13:27application, but it can be mutated whenever you
13:31need to mutate it by an agent.
13:33That means you can customize your whole application just
13:36by asking your agent to change a button
13:39or I never need this thing before.
13:41And it gives you the flexibility to control that UI.
13:46So we start with the basic catalogs
13:49that we used to call them standard.
13:50Now we renamed it to basic to try and make this more clear.
13:53It has a handful of components in it.
13:55You can use them.
13:56They help you get started quickly,
13:58but we expect most of you already have a front end.
14:01You already have your own button component.
14:04We want you to be able to use your button component
14:06and assemble it with all of the rest of your components
14:09to make these widgets.
14:10And you put the widgets together.
14:12They're interactive and you don't have
14:14to use just flat components.
14:16You can use fancy components that are dynamic.
14:18You hydrate them with whatever data sets are available to you.
14:22You define a custom catalog of which are the components
14:25that you're allowed to use.
14:27And that kind of becomes a contract between the front end
14:29and the back end.
14:31Whatever the client advertises, the back end can use.
14:35And I think I'm repeating myself,
14:37but I'm just going to say it again.
14:38You already have a design system.
14:40You already have a front end.
14:42You don't need to change that.
14:44I want you to be able to make your front end.
14:47Maybe it's material design, maybe it's
14:49whatever accessible to your agents and remote agents.
14:55All right.
14:55And so we are largely talking about the web.
14:59I often interact through the web, my laptop.
15:03But let's face it, I'm in meetings all day
15:05and I'm running around and I'm often on my phone,
15:08and then occasionally I'm on a tablet.
15:10And then every once in a while I have something on a TV
15:14and then I'm on my watch.
15:16And like, these are all different types of devices.
15:18And while I love the web, and I have
15:22built most of my career trying to build things for the web,
15:24there are lots of other interfaces out there.
15:27And we want to build a solution that works across all of them.
15:30There are some very, very interesting developments
15:32happening on all of these surfaces,
15:36but I don't have dates and times yet.
15:38But keep watching.
15:39There's a lot of really interesting stuff happening
15:42across Google and also customers.
15:44I've talked to two or three customers that
15:46have bespoke rendering pipelines and bespoke interfaces,
15:49and they're very excited about a standard for declarative UI.
15:52This is not new.
15:53There have been standards for declarative UI before,
15:56but this is a great moment to realize that.
16:00So why do I care so much about the surface.
16:03Security is one of the aspects we're all
16:05familiar with data exfiltration and rendering an image that
16:09allows an agent to accidentally leak data that's
16:11a lot easier to protect against if you're in control of what
16:14the image component is, and if you have structured out
16:17or decoupled the values for that image,
16:20you have layers of defense similarly.
16:26Hidden forms inside of a click target
16:29is another attack vector that you want to protect against.
16:31And the reason, sorry, the way that you
16:33can protect against this is by being
16:35in control of your component library system.
16:39Or we've seen a lot of different supply chain attacks.
16:44One of the ways that you can protect against this
16:47is by being in control of all of the front end
16:49components that you have.
16:50And again, this is on you to control.
16:52But this allows you layers of defense.
16:55And that's really what the security concerns are.
16:57Are defense in depth in UI.
17:01There's a contract between the agent
17:03or all of the agents in the background and the client.
17:06The client is in charge of what components are available.
17:09There's lots and lots of other security concerns,
17:12but know that was one of the main designing principles
17:15in this.
17:15So what you get is a set of core libraries
17:17that handle a lot of the plumbing and the message passing
17:20for you.
17:21Some renderers that we maintain, you're
17:23in charge of your client catalog and pass that in,
17:26and your agent sends the message across
17:27and it shows up for your user.
17:30Where are we at today.
17:32Do we have any transport, any catalog of components.
17:36We do have some tooling that repairs generations if they mess
17:39up a little bit on the fly.
17:41We have a handful of libraries that
17:43are renderers for Web Components or lit.
17:46Angular React.
17:47Flutter these are contributed to by the main owners
17:50of these libraries.
17:52And then we have an ecosystem of other renderers
17:55that have happened around that people are contributing to us.
18:02Today, a-to-I is launched in Gemini, enterprise and Opal
18:06and many other internal tools.
18:08There's lots and lots of teams exploring it.
18:12These demos don't really show the full power.
18:14I'm still pretty excited to get just an input, just a button.
18:18I just want a little bit and I get excited,
18:20but there's so much more that is coming,
18:23as you'll see a little bit later.
18:25We've got some people already exploring generative UI
18:27and coming up with much more exciting demos than this,
18:30but I personally am still very excited to see
18:32this inside of our products.
18:34Showing better information, coming back,
18:37showing a few better affordances,
18:39and rendering fancier components or widgets
18:42whenever we need them.
18:45We also have a few experiments going on, some of which
18:47are public, and we'll make a few more about mixing and matching
18:51between MCP, apps and atooi and some other dynamically
18:54generated UI widgets.
18:57Where else are we going tomorrow.
19:00More integrations and more products.
19:01Again, I'm not.
19:04I cannot leak any of the names or things,
19:06but there are things coming that I'm very excited about.
19:10Stability is a big concern for me.
19:12I really want this to get to a 1.0 release
19:15this year so that we can have long term support
19:18and feel very consistent and stable on it.
19:21I do think that the core infrastructure is now
19:22quite stable.
19:23We're going to add some new features with the release,
19:26and then we're probably going to try for av1 point 1.0,
19:29hopefully.
19:30Q3 and we're going to make it easier to use.
19:34It has largely been get all the parts and pieces right
19:37and make it work.
19:38But we have not yet made it super, super easy for devs.
19:41We made some big strides towards that and the 0.9 release,
19:44but we certainly have more Polish
19:47to do for the developer experience.
19:50And speaking of making things easier for developers,
19:53actually I would like to invite up a Ty to come explain.
19:57Aigoo, aigoo to you.
20:09Hey, everybody.
20:10Thank you.
20:10Alan can everybody hear me.
20:13All right.
20:14Great all right, I'm Ty.
20:18I'm co-founder and CEO of copycat.
20:21For those who don't know us, copycat
20:23makes developer infrastructure for building
20:26AI Copilot, co-pilot, which essentially
20:28means user interactive agents as opposed
20:32to fully autonomous agents.
20:35We make the layers that connect the agentic world
20:39with the user facing application world, and at this point powers
20:43tens of millions of agent user interactions per week
20:47in production from companies like Cisco, Docusign, function
20:50health, Deutsche Telekom and in fact, the majority
20:53of the global 50 as well as the Fortune 500
20:57are today Copilot users.
21:00We're also the company behind the agwi protocol.
21:03If you might heard of it.
21:04It's a protocol that emerged from our partnership
21:07with first chain and then kuii, and has since
21:10been adopted by essentially all the top leaders in the space,
21:13including Google, which has been incredible to collaborate
21:16with in Amazon, Microsoft, Oracle,
21:19and essentially all the top AI framework startups.
21:22So master, pedantic, ag2 and so on,
21:27I'm really excited to be here today and to work in this space.
21:30And the reason is we believe over
21:32the next few years all UI will be
21:35AI, which means that over the next few years,
21:38all interactions between people and technology
21:41are going to become increasingly mediated by agentic systems.
21:45And obviously, this is true of complex software like HubSpot
21:48and Figma and productivity software like Cursor Cloud Code.
21:51But it may even be true of your refrigerator.
21:54And, the first time I heard this,
21:55when we were approached by a large electronics manufacturer,
21:58I thought it was a little bit funny.
21:59But on second thought, wouldn't it
22:01be nice if you could walk by your refrigerator
22:03and tell it to order the ingredients that are missing
22:06to make lasagna tonight.
22:08So that's the world we're going to.
22:10But where are we today.
22:12Many of you audience probably know
22:14that building great user interactive agents is very
22:19difficult. And the core reason is
22:22that agents break the request response paradigm that's
22:25been powering the internet for the last 30 years, right.
22:28So agents are, of course, just software.
22:30But from the standpoint of software,
22:32they're a little bit like a weird bird.
22:34They have their long running, so they
22:36have to stream their work over time and support reconnections
22:40and even mid-run steerings.
22:42They have to support structured and unstructured Data
22:45Exchange at the same time.
22:46So they have tool calls and state updates at the same time.
22:49You have text and voice.
22:50There's really a long list of these peculiarities
22:52and difficulties which make it challenging to build
22:55great generic applications.
22:57And that's where Copilot and ag come in.
23:00So Copilot provides open source developer SDKs
23:04and a self-hostable Cloud for building generic applications
23:08in a way that's optimized for the generic paradigm,
23:11as opposed to retrofitted from request response
23:14in AG, which again stands for the agent user interaction
23:17protocol lets developers who build agentic frameworks.
23:20And harnesses bring their generic backends
23:22ends to any custom hygienic front end
23:25by building to a robust standard.
23:26So it makes it very easy, very simple to conform
23:28to the requirements and everything just
23:30works and continues to automatically
23:32evolve with the ecosystem and to situate it in the landscape.
23:35You obviously have MCP connecting agents
23:37to tools into context and to connecting agents
23:40to other genetic systems.
23:42AGI is the third leg of this triangle connects your agents
23:45to user facing services.
23:49Our initial focus for AGI has been the web,
23:51but as Alan mentioned, there's quickly
23:53growing list of front services supported by AGI,
23:56including Slack, WhatsApp, mobile and other services.
24:03Here's an architectural overview of AGI.
24:05We actually don't have time to get into this,
24:07so you want to take a picture and ask me questions about it
24:09afterwards, I'll leave it for a moment.
24:14All right.
24:16What's become very clear to us, working
24:19with all these top leaders over the past couple of years
24:21is that no single generative UI solution
24:24is good for all different use cases,
24:26and generative UI solutions really
24:28live on a spectrum that extends from more
24:31control to fully open ended.
24:34And AGI supports this full spectrum,
24:37which means you can bring the right tool for the right job
24:40at any given time.
24:42And now I'm going to speed through the three key pillars
24:45of the generative UI spectrum controlled, declarative and open
24:49ended generative UI.
24:51So first control of UI wear developers
24:54define a collection of pre-built components
24:56which the agent can call upon when he wants to show the UI.
24:59So first, let's just see it in action.
25:03So here we see a user asking to see revenue split by category,
25:08and instead of getting kind of an impenetrable long paragraph
25:11of text, you see this beautiful pie
25:13chart that is much more easy to understand from a code
25:18standpoint.
25:19It's extremely straightforward.
25:20Here, for example, we have the use component hook in React
25:22where you define the component and what
25:24it's for the parameters.
25:26Whoops the parameters that it requires
25:30in just a React renderer.
25:31In this case a pie chart.
25:34To some extent controlled generative
25:36UI is the most boring part of this spectrum,
25:39but it does give you Pixel perfect designs
25:42and maximal determinism, which makes
25:44it a really great fit for the few most used surfaces
25:47in your applications.
25:48For example, if you're an airline,
25:49you want your flight cards to be Pixel perfect and as
25:54deterministic as possible, and for that reason,
25:57we see control UI is the workhorse of the generative UI
26:00spectrum.
26:02Next, we have declarative UI where UI fits in
26:06and where developers declare a component catalog of LEGO
26:10like building blocks that the agent then assembles on demand
26:15to answer any user queries.
26:17So let's see that in action.
26:22All right.
26:23So here we see a user asking for a sales dashboard that includes
26:26revenue and customer churn.
26:28And it includes all the information
26:30a salesperson might want to see instead of again
26:32getting a block of text.
26:33You see this beautiful sales dashboard
26:34that combines cards with titles and pie charts and bar charts.
26:38Each one of these components individually
26:40was predefined by the programmer,
26:42but the collection of them together as a sales dashboard
26:46was assembled by the agent on demand.
26:49And to look at some code to define this, you first
26:51define the catalog definitions, what each component is
26:55and what arguments it acquires from the outside.
26:57You'd need a catalog renders to render those this data in React
27:01in this case.
27:02And that's it right things do the AG A2 UI handshake.
27:07That's a mouthful.
27:09That's all you need to do to get any AGI agent, which is again,
27:12essentially all of them to speak to UI to your users
27:15in your own custom application e.g.
27:18middleware absorbs all the additional complexity,
27:22and with declarative UI you're giving up
27:25some Pixel perfect perfection.
27:26You define the components, but every component
27:29has to be an assembly of these pre-built building blocks.
27:32And you're giving up some determinism
27:33because the agent is assembling these components together,
27:36typically.
27:37But what you get in return is very broad support
27:40for different user queries with a one time definition
27:43of a component catalog.
27:44And this profile trade offs makes it a really great choice
27:47for the long tail of user facing services
27:50in consumer applications, and also for internal enterprise
27:54applications where the efficiency of implementation
27:57is really important.
28:00And finally, at the far end of the spectrum,
28:02we have open ended UI.
28:04So in the previous two approaches,
28:06we at build time prescribe what types of UIs
28:10the agent can show with open ended generative UI,
28:14you're opening the door for the agent
28:15to display anything at all.
28:17And typically that's achieved via an embedded iframe.
28:19And there are two kinds of open ended UIs
28:21that are worth calling out.
28:22Of course, MCP apps that I showed us before.
28:24They're really optimized for third party tools
28:27that you want to bring into your own application.
28:29And again, there's a large collection
28:30of these that are built for the Super host.
28:32For ChatGPT for Cloud for cursor and so on.
28:35And then the second kind is fully open
28:37ended generative UI, where the agent actually
28:39generates fully custom HTML live inside
28:42of a secure sandbox for security purposes,
28:45but they can show really anything at all.
28:47So we already saw a demo of MCP apps.
28:49So here's a demo of fully open ended generative
28:51UI where the user asks for a specialized calculator
28:54for some specific use case.
28:55They have even a specific design that they want,
28:58and the agent is building it on the fly.
29:00And as you can see, it's a little bit slower.
29:02It's less deterministic, but you do
29:03get at the very end and actually fully functioning component
29:07that nobody could ever think to design ahead of time.
29:12And again, with AG Aggie MCP, apps and open
29:16ended generative UI are just a configuration away
29:19in your own custom full stack agents,
29:22so you can pass the URL of any MCP apps
29:24you want to use and just turn on the open generative UI flag.
29:27And now any of your backend agents
29:29without any further customization in the backend
29:32can support these experiences.
29:35So to sum up, we covered the three pillars
29:38of the generative UI spectrum, starting
29:40with controlled generative UI where
29:43you define pre-built components that are good for your few most
29:47used surfaces.
29:49Then declarative generative UI of LEGO
29:51like building blocks that are assembled together,
29:53which are great for the long tail of UIs
29:55and for internal enterprise applications.
29:57And finally, fully open ended UIs,
30:00which are really great for third party applications
30:02and for fully custom experiences.
30:06If you'd like to speak about this,
30:08you can email me while we're all in town.
30:10My email is atai at.
30:11Capecchi also be outside here answering questions,
30:14doing some impromptu demos.
30:15And we actually have a special offering
30:17for folks at Google Cloud Next to help your engineers ship
30:22your applications to production so you can
30:24scan the QR code if you'd like.
30:27And now I'm really excited to bring Nico from Akka to show us
30:31how all of this theory is put to practice.
30:40Thanks, sati.
30:41Hi, everybody.
30:42So now I'm going to talk to you about why
30:44brands and big enterprises needs to get into January.
30:50I'm Nicolas, I'm the CTO EMEA at akaaka.
30:55I'm part of WTP.
30:57Why it's important.
30:59There is a structural shift.
31:00For the past 20 decades, we have built
31:03websites which were optimized around navigation.
31:07We take it for granted, right.
31:09You navigate browse, Bryce.
31:10You go to static pages.
31:12And from that, companies were inferring your intent,
31:17and we could personalize experiences.
31:19But now this is changing.
31:22You're doing the discovery of products and services
31:25already in ChatGPT in Gemini.
31:28And when you get onto a website, you already have a clear goal.
31:31And either you find the answer from this goal on the page,
31:35or you will move back to ChatGPT and Gemini
31:37to continue your searches.
31:39So navigation is not enough anymore.
31:41We need solutions that can adapt.
31:44That takes us to the content economy.
31:47First of all, it's built intent is the primary signal.
31:53Secondly, you need systems which are adaptive, which assemble,
31:56which are not fixed.
31:58And the last one is changing the trust
32:01between the system and user.
32:03Because now we are actually capturing real intent.
32:07We're responding to an intent.
32:08We're not trying to track you and infer what you want.
32:14So if intent becomes the primary signals,
32:16there is a fundamental challenge for enterprises.
32:19Because human intent is rich, it's messy.
32:22It's nuanced.
32:24Enterprises can't work with messy elements like that.
32:27So we need structured intent.
32:29How do we do that.
32:31Well, we go back to work that was done in the 90s as the jobs
32:36to be done framework.
32:37And it tells us that human intent is actually
32:41built on three elements.
32:43The functional part.
32:44What are you looking for.
32:46The size, the color of the object you're looking at.
32:49But there's also the part which is super important.
32:53What is the confidence level.
32:55What is the risk level.
32:56What is the urgency of your need.
32:59Very importantly, the social part.
33:01How do you want to be perceived.
33:04The politics at play.
33:05The pressure.
33:06So real decisions are made with intent.
33:10Object functional.
33:11Emotional and social.
33:13So we're going to look at how we do that with a real world
33:16scenario with a company called synthesis.
33:20It's a fraud detection platform which
33:23helps banks, fintech and e-commerce e-commerce
33:26detect fraud at payment level.
33:29It's a fictional brand, but it's built on real work.
33:32I can't share the real work because it's confidential,
33:35but we rebuilt the demo with the synthesis.
33:40So now I'm going to show you actually
33:42in the backend how we're capturing
33:44the intent of a future consumer.
33:48So here, it's taking the intent from a new security
33:53officer in a startup, which the recent rises
33:56in cross-border payment.
33:58And what we're doing here is our Gemini agent
34:01is splitting that intent into functional, personal, Lord,
34:05personal and emotional.
34:06From a functional point of view, it's
34:08telling us that person is looking from a fraud detection
34:11solution, emotional point of view.
34:13That person needs to feel competent in his choice
34:16and from a social element.
34:18Very importantly, we assume leadership pressure for him
34:21to make this choice.
34:22This solution is creating an intent object
34:27that gives us product matching the tone of voice
34:30that we need to use to talk to that user.
34:33And the experience design we need to create.
34:36So how do we go from intent to interface now.
34:40Well, three elements.
34:42The interpretation layer.
34:44We've just seen it.
34:44We've taken intent.
34:46We break it down.
34:47Then we pass that to a brain.
34:49Brain this is where all of the product data, the brain data
34:52is stored.
34:53And it also has governance and compliance.
34:56And finally, we pass this to the generative interface
35:00which takes this content and assemble it on the fly.
35:04The separation is very important because we
35:06want enterprises to be able to control this system.
35:10Let's look at the brand right now.
35:12So brand brand at the core has a vector database which
35:15has all of the product data.
35:17But we also have a knowledge graph here,
35:19which has all the complex enterprise context.
35:22And finally, we have a layer dedicated
35:24to compliance legal compliance for from the company.
35:30And the separation of content is very important again.
35:33So we give control to companies right.
35:35We don't want to be an agent in a black box
35:37where everything is controlled in prompt.
35:40The output of the brain drain is a validated content object
35:43that is given to the generative interface layer.
35:47So it takes that intent structured content.
35:50Then it will select the components from the design
35:53system and assemble the layout and inject the content into it.
35:59This is then always done within the design system constraint.
36:03So let's look at a first demo of that website.
36:07So here this is the synthesis homepage.
36:11You can see standard homepage talking
36:12about the actual product.
36:15The USBs.
36:17But at the top of that page we allow
36:20that visitor to enter his real question his intent.
36:27And here it's someone that wants to improve fraud detection
36:32without hurting conversion.
36:35So it's giving it that intent.
36:37The system we're going through all the layers the brain,
36:40brain, the interpretation layer.
36:42And we actually generating a completely new page
36:47with the right component to answer his needs.
36:49So here we have generated a full solution pages
36:54with our reassurances.
36:56All of the content that is here has been rewritten by the brain,
37:00but the components is used from the design system.
37:06It has, as you can see, some Roi which
37:09has been picked specifically based on that intent.
37:13Some specific content about fraud detection.
37:17Pull that quote that makes sense based on the user search.
37:24So let's now look at a different intent from a different user.
37:28In this case, it's someone which is
37:30looking to fix fraud spike extremely fast.
37:33They've seen fraud happening and they
37:35want something that fix it in weeks, not months.
37:39So the system gets that intent and will build the page.
37:44At this point, the system understood the urgency, right.
37:47The social and emotional aspect of that.
37:50And it's been on the page with a decent system.
37:52But here it's a much more focused response.
37:55It's not full detailed page.
37:58The pages is created to go straight to the next best
38:02action for that consumer.
38:07So for that team that we work with, what it means
38:11is these type of responses.
38:14They used to do it by hand create.
38:16They created PDF created by humans.
38:19It used to take them more than six hours.
38:22Now, this is done in tens directly on their website.
38:26So massive acceleration.
38:27They've also clearly counted how much money
38:32they were spending creating this dedicated response.
38:35And that cost obviously now is gone.
38:38And more importantly, they're telling us
38:40that we're no longer spending hours preparing responses.
38:43We can focus on the strategy discussion from the start.
38:46It's really changing how this company
38:50is engaging with its leads.
38:54So now content and is infrastructure
38:58for company navigation is replaced by intent.
39:03Static pages needs to become dynamic system.
39:06So as enterprises for the next web experiences,
39:11you need systems that understand reason and Compose in real time.
39:18And I will give it back to Alan.
39:20Thank you.
39:29It has been a privilege and a joy
39:31to work with the other presenters
39:35that you saw today, with all of the engineers that
39:39have got us here.
39:40And honestly, with all of you, we
39:42are working on things that are new
39:45and we're coming up with really good foundational technologies
39:48and capabilities ourselves, but we're doing it together
39:51to build the new versions of the web.
39:57I'm putting this QR code here and on this
40:01thank you slide at the end.
40:02Same QR code to a form.
40:05It's an anonymous form.
40:06You can put in contact information if you want to.
40:11This is what would you like from generative UI projects in 2026.
40:17We are working together at least monthly, often
40:19much more frequently.
40:21We will continue to do so this year.
40:23What would you like to see from any of these projects.
40:27Please tell us.
40:28We'll also be hanging out in the hallway or whatever,
40:31and you can come find us here.
40:32But I do want to hear from you.
40:34We are solving our problems and the problems that we see
40:37and the problems that we hear from you,
40:39but this is an opportunity.
40:41Please tell us what you want.
40:43We are all in this together.
40:45And thank you for joining me.