Google’s A2UI Express aims to make generative interfaces faster

Google has introduced A2UI Express, a new way for AI agents to describe app interfaces that the project says can reduce generation latency by up to three times. Developers can also use pre-defined layouts and decision models for suitable screens, trading some flexibility for much faster results.

Watch Desk analysis

What happened

A2UI is Google’s protocol for agents to create interfaces from components that an app has explicitly made available. Until now, agents generated the interface description as JSON. Google says that can be slow and error-prone, so its new Express format uses a more compact, Python-like syntax that the SDK compiles into standard A2UI messages.

Google says Express uses up to three times fewer tokens than Direct JSON for similar interfaces, and can reduce latency by up to three times. For screens with a known structure, developers can instead use a decision model such as Jev to select from layouts defined in advance. Google cites an Apparite demonstration in which the component tree appeared in 92 to 298 milliseconds; the complete screens took 0.8 to 1.7 seconds. That demonstration covered 11 prompts, so it is an early example, not a broad performance benchmark. Google’s A2UI announcement explains the formats and setup.

Express is available in Google’s Python and TypeScript agent SDKs, with a Dart SDK version also listed. Developers can swap inference formats without replacing their component catalogue or renderer.

Why it matters

The change gives developers a practical choice: let a model build a flexible interface with Express, or have a faster decision model choose among layouts already designed by the team. In the latter case, the model has less freedom, but the interface can appear sooner. That makes the trade-off more useful than a single headline speed claim suggests.

Our read

This is a substantial improvement to the plumbing behind AI-generated interfaces, with a clear decision for developers rather than a vague promise that everything will feel faster. Try Express when screens need flexibility; use controlled layouts when the shape is known and speed matters more. Google’s headline figures are worth attention, but the small demonstration is not a substitute for testing on your own models and screens.

What to watch

• Whether developers see similar latency gains in production, across different models and catalogues.
• How often known layouts can cover real screens without making the experience feel rigid.
• Whether the SDKs’ format-swapping approach attracts other inference formats and tooling.

Discussion spark: For AI-built interfaces, would you favour the speed of pre-defined layouts or the flexibility of letting a model build the screen?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.