# Google's A2UI adds swappable inference formats and a faster Express default

published: 2026-10-27T16:05:29.126Z
topic: frameworks
source: https://developers.googleblog.com/fast-reliable-a2ui/
url: https://aiagentnews.io/stories/google-s-a2ui-adds-swappable-inference-formats-and-a-faster-express-default

Google's A2UI team has released swappable inference formats for its generative UI protocol, with A2UI Express as the new default. The company says Express can cut latency up to 3x and that decision models such as Jev can cut it up to 80x for layout choices.

A2UI lets an agent describe a user interface built from a client's trusted component catalog, and the client's renderer draws it. Until now agents generated the standard A2UI JSON directly, which the team found slow and error-prone for LLMs. The new A2UI Express format is a compact intermediate format that the Agent SDKs compile into standard A2UI JSON, and the team reports up to 3x fewer tokens than Direct JSON for the same UI.

Developers pick a format with one argument when creating the generator, and the SDK produces both the prompt instructions and the compiler for it. Express is available in the Python agent SDK (a2ui-agent-sdk 0.9.0), the TypeScript package @a2ui/agent and the Dart SDK. In tests across four formats on a 51-task eval with gemini-3.5-flash, Express performed as well with thinking off.

For screens with a known shape, a System One decision model such as Jev, from TypeSafe, can choose among predefined layouts instead of the model designing one. Apparite (jev2ui), an open-source design tool, uses this approach, and in one run of 11 prompts its component tree appeared in 92 to 298 ms. The project describes these figures as a first look because the sample is small.

## Why it matters if you build agents

Agents can trade expressivity for speed: Express suits screens that vary, while a decision model suits screens with fixed layouts defined in advance. Custom formats can be added by supplying a prompt generator and a converter, without changing the catalog or renderers. Controlled generation only produces the layouts its grammar describes, so unfamiliar screens still need Express or another format.