---
type: "post"
id: "fast-reliable-a2ui-with-swappable-inference-formats-f5e911"
slug: "fast-reliable-a2ui-with-swappable-inference-formats-f5e911"
title: "Fast, reliable A2UI, with swappable inference formats"
description: "Google has added swappable inference formats to A2UI and made A2UI Express its new default format."
retrieval_nugget: "Google has added swappable inference formats to A2UI and made A2UI Express its new default format. Instead of asking a model to emit a full JSON interface tree, teams can use compact model-facing syntax compiled by the SDK into validated A2UI messages, targeting lower latency and fewer generated tokens."
published_at: "2026-10-28"
updated_at: "2026-10-28"
record_date: "2026-10-28"
date_kind: "published_at"
topics: ["google","gemini"]
entities: ["Google","A2UI","A2UI Express","Jev","TypeSafe","Gemini 3.5 Flash"]
related_patterns: ["agent-protocols-become-interoperability-layer","model-answers-become-task-interfaces","agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "Google introduces A2UI Express and swappable inference formats, claiming up to 3× lower latency and up to 80× with decision models—distinct from the archive’s general inference coverage."
prior_refs: ["memory:659ae474-e509-8f77-a64e-cd53c5e8d8c2","memory:3dfd874a-49b6-8bc0-a2cb-9cc0ca3cf2d1"]
source_url: "https://developers.googleblog.com/fast-reliable-a2ui"
source_urls: ["https://developers.googleblog.com/fast-reliable-a2ui"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:fast-reliable-a2ui-with-swappable-inference-formats-f5e911"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/fast-reliable-a2ui-with-swappable-inference-formats-f5e911/","markdown":"https://newruntime.com/posts/fast-reliable-a2ui-with-swappable-inference-formats-f5e911.md","json":"https://newruntime.com/posts/fast-reliable-a2ui-with-swappable-inference-formats-f5e911.json"}
---

# Fast, reliable A2UI, with swappable inference formats

## Retrieval answer

Google has added swappable inference formats to A2UI and made A2UI Express its new default format. Instead of asking a model to emit a full JSON interface tree, teams can use compact model-facing syntax compiled by the SDK into validated A2UI messages, targeting lower latency and fewer generated tokens.

## Summary

Google has added swappable inference formats to A2UI and made A2UI Express its new default format. Instead of asking a model to emit a full JSON interface tree, teams can use compact model-facing syntax compiled by the SDK into validated A2UI messages, targeting lower latency and fewer generated tokens.

## What happened

Google’s A2UI team has released an update aimed at a practical bottleneck in generative interfaces: waiting for a model to generate each screen. A2UI is an agent-oriented protocol in which a client advertises a catalog of trusted components. An agent then sends A2UI messages describing a custom interface assembled from that catalog, and the client renderer draws it. The catalog is the contract: the agent is meant to compose available components rather than invent arbitrary UI primitives.

Initially, agents generated the JSON transport messages directly. That means generating every brace, nested object, and property in a tree constrained by a complex JSON schema. Google says customer feedback exposed latency and accuracy problems with that approach. It also coupled A2UI closely to text-generating language models, making it harder to use faster decision-oriented models. The new abstraction is an `inference_format`: the format a model generates before SDK code turns it into standard A2UI transport messages.

An inference format has two schema-derived parts: prompt instructions that teach its syntax and a compiler that converts a model reply into validated A2UI messages. A builder selects the format when creating an A2UI generator; the SDK supplies the relevant prompt snippet and parses the reply into text parts plus validated messages. Google says this design prevents a format from describing components or properties absent from the catalog. Builders can use A2UI Express, or provide another domain-specific language and compiler for a specialized case, while retaining compatibility with A2UI transports and renderers.

A2UI Express is the new default. It represents a screen as one Python-like statement per line, with component variables, positional arguments, bindings to a data model, and declared events. Google says Express needs up to three times fewer tokens than Direct JSON for the same kinds of UI and can reduce latency by up to three times. It is available in `a2ui-agent-sdk` 0.9.0 for Python, `@a2ui/agent` for TypeScript, and `a2ui_agent` 0.0.1-wip004 for Dart. Python and TypeScript can stream Direct JSON chunk by chunk; Express is compiled from the complete reply.

Google also describes a more constrained path for screens whose layout is already known, such as settings, checkout, or confirmation pages. Its example uses TypeSafe’s Jev, a System One decision model, to answer typed layout questions with calibrated probabilities. Google says Jev can answer many questions in one request in roughly 100 to 300 milliseconds. In the cited `jev2ui` design tool, Jev selects screen structure, code constructs A2UI components, and Gemini 3.5 Flash Lite writes only the text. The layout questions and their criteria are authored beforehand in a Markdown grammar.

## What the archive adds

We have not covered this before; it opens a new thread.

## Trend signal

This supports [Agent protocols become the interoperability layer](https://newruntime.com/patterns/agent-protocols-become-interoperability-layer/). A2UI keeps a stable transport and renderer contract while allowing the model-facing representation to change. It also supports the archive assumption that platforms can preserve organization-defined evaluation and integration surfaces while models are swapped, as described in [Owned learning loop](https://newruntime.com/shifts/owned-learning-loop/).

## Why now

Google frames the release as a response to feedback on direct JSON generation: complex schema-constrained trees were slow and error-prone. A2UI initially launched with Direct JSON in December 2025. The new format layer addresses that limitation without requiring a change to the trusted component catalog or the frontend renderer.

## Why it matters

The relevant decision for a product team is no longer simply which model generates an interface. It is which representation the model should generate, and which portions of a screen need language generation at all. A compact syntax can reduce output volume and avoid asking a model to reliably reproduce a deeply nested serialization format. A constrained decision path can go further when the set of layouts is known.

The boundary also has operational value. The catalog remains the allowed UI surface, and the SDK compiler produces validated A2UI messages. That lets teams experiment with model families or specialized DSLs without rebuilding the client-side integration each time. It does not eliminate the need to test outputs, but it moves some correctness work from probabilistic text generation into deterministic translation and schema validation.

## What happens next

- **More A2UI implementations will add task-specific inference formats beyond Express.** (6 months, 65%). Watch for: SDK format factories and published custom-format examples.
- **Known-layout flows will increasingly split layout selection from copy generation.** (12 months, 60%). Watch for: decision-model integrations and grammar-based UI definitions.

## Limits and uncertainty

The reported “up to” threefold latency and token reductions are Google’s claims, not a universal benchmark. The post mentions a 51-task evaluation on Gemini 3.5 Flash with thinking on and off, but the supplied material does not provide the task set, full scores, token totals, latency distributions, or comparisons across other models. Express also compiles only after a complete reply, unlike streamed Direct JSON in the Python and TypeScript SDKs; the end-user effect will depend on an application’s streaming design. The Jev example illustrates an architecture rather than proving that it is best for every interface.

## Skill to build now

Choose one existing agent-driven screen and write its component catalog as an explicit contract: allowed components, properties, bindings, and actions. Generate the same screen through Direct JSON and Express, then measure end-to-end time, generated-token count, validation failures, and perceived time to first useful UI. For a fixed-layout flow, separate layout choices from generated copy and list the decisions that can be represented as typed questions before introducing a model.

## Reader takeaway

A2UI’s update makes the model’s output syntax replaceable while preserving the UI protocol beneath it. For builders, that creates a practical performance lever: keep the trusted renderer contract stable, then choose a compact generative format or a constrained decision system according to how much of the interface is genuinely open-ended.

## Open question

Will custom inference formats remain portable across models and SDK languages, or will their prompts, compilers, and evaluation behavior become a new source of platform-specific lock-in?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)