# A2UI Express Explained: Faster Generative UI Without Replacing the Renderer

Google's compact inference format reduces the text an AI model has to generate, but faster completion and faster first paint are different goals.

By Paras Madan | 2026-10-28


Google announced A2UI Express and swappable inference formats on October 27. Models can describe interfaces more compactly, while the SDK converts their output into standard A2UI messages for existing renderers.

The useful change is a separation between what the model writes and what the application consumes. A model does not necessarily need to generate every structural detail of the final message.

## What A2UI Express changes

A2UI lets an agent describe interfaces using a catalog of components the application supports. With Express, the model writes compact component declarations rather than the full JSON representation.

The [Express specification](https://a2ui.org/specification/proposals/express/a2ui_express/) describes variable assignments, component references and bindings to application data. A compiler maps the compact arguments to the catalog's schema and assembles the protocol messages. The generated format resembles Python, but it is a declarative interface description, not arbitrary Python that the application should execute.

For example, a notification card needs text, an icon and actions. Repeated JSON property names do not add meaning to those choices. Moving that repetitive structure into a compiler gives the model less to write while keeping the application's component contract explicit.

**A valid component tree is not proof that the interface makes the right decision.** A correctly formatted confirmation screen can still show an unsuitable action or misleading text. Product permissions and business rules remain application responsibilities.

## The latency claim needs a streaming caveat

In its [announcement](https://developers.googleblog.com/fast-reliable-a2ui/), Google reports up to threefold reductions in tokens and latency, with format comparisons using 51 tasks on Gemini 3.5 Flash. Those are Google's results, not a universal speedup for every catalog or model.

There is an important implementation detail: the announcement says Python and TypeScript can stream Direct JSON incrementally, while Express is compiled from a complete reply.

Test time to the first useful screen element separately from completion time. Less output does not automatically mean a user can begin reading sooner.

## SDK availability and migration

Express is available in Python, TypeScript and Dart agent SDKs, but October 27 was not their shared release date.

The [Python SDK changelog](https://github.com/a2ui-project/a2ui/blob/main/python/a2ui_agent/CHANGELOG.md) dates version 0.9.0 to October 17. It includes the move of Express out of its experimental import path and several breaking changes, including dependency and package restructuring.

An existing integration should therefore pin its versions and review migration notes before changing the inference format. Keeping the renderer unchanged does not mean every server-side import remains compatible.

## When a known layout should stay a known layout

Google also points to [Apparite's Jev and A2UI example](https://github.com/dglazkov/jev2ui), which separates structural decisions from text generation. Code assembles the component tree from those decisions while a language model supplies the wording. Its comparison includes a baseline that asks Gemini to generate A2UI directly.

This is a separate design pattern, not evidence that Express itself delivers the announcement's larger decision-model speedup. The example also requires both Jev and Gemini credentials, so its operating cost cannot be inferred from a token reduction alone.

For a familiar settings page, repeatedly inventing a layout may add little value. For an interface whose structure genuinely depends on the request, more expressive generation can be useful. The engineering decision is how much uncertainty the model needs to resolve before the user can do something useful.

The next optimization may come from asking the model to make fewer decisions, rather than asking it to write the same screen faster.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain developer tools and AI engineering through clear writing and launch stories people can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min)