twimco.Let’s talk ↗︎
Menu
Field notes / Product & engineering

Google's A2UI 1.0 lets an agent draw screens inside your app. Sort your screens into three bins before it draws any

Ask your app's assistant to compare three plans and today it answers with a paragraph. With A2UI 1.0, it could answer with a real comparison screen built from your own components. That's a better answer, as long as you've decided which screens an agent should never compose.

Editorial engraving of a skylit product room where a designer, an engineer and a compliance lead sort printed phone-screen cards into three trays.

A2UI 1.0, released by Google on November 11, 2026, is a stable open protocol that lets an AI agent describe a screen, such as a form or a comparison, which your app then draws with its own native components. Renderers exist for the web, Flutter and iOS, with Android's Compose renderer still catching up. Before adopting it, sort your screens: ones an agent may compose, ones it may propose for the user to confirm, and ones that stay hand-built.

What Google released

On November 11, Google released A2UI 1.0, the first stable version of its open protocol for what it calls generative UI. The idea fits in one sentence from Google's post: the agent describes what to show, and every app draws the UI with its own native components.

Today, if a customer asks your app's assistant to compare three plans, they get a paragraph. A2UI lets the agent answer with a screen instead: a comparison, a form to confirm, a chart. The agent sends a description in a structured format. Your app's renderer turns that description into real buttons, lists and fields from your own design system. Android's documentation for its renderer stresses that it renders native components without executing arbitrary code.

Version 1.0 adds bidirectional function calls, so a picker for flights or reservations can check availability while the user is choosing, plus support for accessibility and extension metadata. Google lists renderers for React, Angular and Lit on the web, Flutter across platforms, SwiftUI on iOS and Jetpack Compose on Android, and agent SDKs for Python, TypeScript and Dart. Atlassian, Alation, Bloomberg and ByteDance are named as building with it. The web and Flutter renderers also work inside AI assistants that host MCP Apps, which Google says includes Claude and Gemini.

Why it matters for a mobile app

A mobile screen is small and a paragraph is a bad answer on it. If your app already has an assistant, or you're planning one, A2UI is a plausible way to make its answers look and behave like the rest of the app instead of a chat bubble bolted on.

It also changes who designs a screen. Until now, every screen in your app was drawn by a designer, built by an engineer and reviewed before release. With A2UI, some screens will be assembled at runtime by an agent, from parts you've provided, in combinations nobody reviewed.

That's fine for some screens and not for others. The agent decides what to show. Your app decides what it is allowed to show. That second decision is the one to make before you write any code.

Decision guide

Three bins for every screen

ConsiderationExample screensWhat the agent may do
ComposeOrder status, plan comparison, store hours, a chart of past bookingsDraw it from your components with read-only data
ProposeRebook an appointment, change a delivery slot, fill a return formPre-fill it; the user confirms on a screen your app controls
Hand-built onlyPayment, consent, identity checks, account deletion, anything legalLink to the existing screen; never compose it
Decision framework for a product team; illustrative examples, not A2UI guidance.
The agent decides what to show. Your app decides what it is allowed to show.

Sort every screen into three bins

Take the list of screens in your app, or the list of things customers ask your support team about, and put each one in a bin.

Compose. Screens that only show information the user is already entitled to see, and where a slightly odd layout does no harm. Order status. A comparison of plans or products. Store hours and locations. A chart of past bookings. Let the agent build these from your components, with read-only data.

Propose. Screens that change something, but where the user can review it before it happens. Rebooking an appointment. Changing a delivery slot. Starting a return. The agent can pre-fill the form and show it; the confirm step should be a screen your app owns, with the same wording and checks as anywhere else in the app.

Hand-built only. Screens where the exact wording, order and checks matter legally or financially. Payment and card entry. Consent and privacy choices. Identity verification. Account deletion. Anything a regulator or a lawyer has looked at. The agent can link to these. It should never compose them.

Most teams find the third bin is longer than they expected, and that's the point of doing the exercise. Our note on why the admin screen is where your software earns its keep makes a related case: the screens with consequences deserve the most design attention, not the least.

What to check before you plan

Renderer maturity. Google's post calls the protocol stable. The renderers aren't all there yet. According to a2ui.org's renderer table, the web renderers are stable for 1.0, while Flutter and SwiftUI are listed as alpha. Android's Jetpack Compose renderer supports version 0.9.1 today, with 1.0 coming soon; the androidx.a2ui library was at 1.0.0-alpha02 as of October 21. If your app is native Android, plan on that gap.

Your component catalog. The agent can only draw what your renderer exposes. Decide which components are in the catalog. Leaving the payment form out of it is the simplest guarantee that the agent can't compose one.

Accessibility. Version 1.0 adds accessibility support to the protocol. Test generated screens with screen readers and large text the same way you test hand-built ones. A screen nobody designed still has to work for everyone.

Upgrades. Version 1.0 isn't a drop-in for anything you prototyped earlier. The specification renames the message envelopes and changes how bindings are written, and Google points 0.9 users to an evolution guide that covers every change. If a vendor demo was built on an older version, ask which one before you plan around it.

Testing. Hand-built screens get tested before release. Composed screens can't be tested exhaustively, because nobody knows every combination in advance. What you can test is the catalog: each component on its own, with long text, missing data and a slow connection, and a set of typical requests run against the agent before each release.

Logging. When a customer says "the app showed me the wrong price," you need to know what the agent sent and what the app drew. Log the description, not just the conversation.

The assistant outside your app. Because A2UI screens can appear in AI assistants that host MCP Apps, your screens may show up somewhere other than your app. Decide whether you want that before someone asks.

Our note on Siri taking actions inside any app works through a related question for Apple's platform: what an outside assistant should be able to do in your app.

Sequence of work

Where the renderers stand

  1. Web

    React, Angular and Lit renderers listed as stable for 1.0.

  2. Flutter

    Mobile, web and desktop; listed as alpha.

  3. iOS

    SwiftUI renderer for 1.0, listed as alpha.

  4. Android

    Jetpack Compose renderer supports 0.9.1 today; 1.0 is coming soon (androidx.a2ui 1.0.0-alpha02, Oct 21).

Renderer status per sources read November 13, 2026: Google's November 11 post, a2ui.org and Android Developers. Check again before you plan.

Before the first generated screen

Run the three-bin exercise with design, engineering and whoever owns compliance in the same room. Turn the result into the component catalog your renderer exposes, so the rules live in code, not in a prompt. Start with one composed screen, like order status, and watch it in production before you add a proposed one.

If you're weighing whether an agent belongs in your app at all, or how to build one that works with one hand, a bad connection and an interruption, that's our mobile application development work. Talk to us.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎