AI interfaces are starting to outgrow the chatbox. Instead of returning only text, agents can assemble the interface a user needs in real time. The challenge is enabling that flexibility without giving an AI model unrestricted control over the frontend.

This talk presents a bounded architecture for generative UI using React and Google’s A2UI protocol. We’ll explore how agents can stream declarative interface schemas that are validated and mapped onto trusted components from an existing design system. We’ll also cover component contracts, server actions, client-side state, streaming, and the boundary between agent-generated structure and application-controlled behaviour.

Client-side navigation has always been the SPA advantage over server-rendered apps - click a link, see the next page instantly, no round trip. Next.js has been closing that gap piece by piece, and its latest rendering model gets it most of the way there without giving up streaming, caching, or Server Components. This talk breaks down how Next.js decides what's static and what's dynamic at the component level rather than the route level, why that shift is what actually makes instant navigation possible, and what's involved in bringing it into an existing production app.

Most teams hear "AI design system" and reach for the thing they already know: a component library, now with a chat panel bolted onto it. I think that's the wrong shape. A design system in an AI-assisted codebase isn't a catalogue of components; it's a set of architectural contracts that a coding agent can actually be held to. Tokens, composition rules, and constraints that are legible to a model, not just to a designer in Figma.

The talk walks through what that looks like in practice: what an agent gets wrong when your system is only documented in prose, how to encode intent so the generated code is boringly consistent instead of plausibly wrong, and where the boundary sits between what you let a model generate and what you pin down by hand.

What attendees can expect to take away:

• A working definition of an AI design system that isn't just "components plus autocomplete"
• The failure modes I've hit shipping React/TypeScript UI with coding agents, with real examples
• Concrete techniques for encoding design constraints so agents produce consistent output
• A sense of what to fix first if their current system is fighting their tooling rather than helping it

It's aimed at frontend and product engineers, and at design-system maintainers wondering what to do about all this. No prior AI-tooling experience needed.