Google A2UI v0.9 lets AI agents render native UI without code

Key Takeaways

• A2UI v0.9 shifts from agents generating arbitrary UI code to declaring intent that apps render natively
• New Python SDK handles version negotiation, streaming, and partial LLM output healing
• Works over MCP, WebSockets, REST, AG-UI, and the new A2A 1.0 transport layer

Google has released A2UI v0.9, an open standard that lets AI agents request user interfaces without generating executable code. Instead of shipping components, agents declare what they need. The host app renders it using whatever design system it already runs.

The release, announced on the Google Developers Blog, marks a philosophical shift. Prior approaches let models output raw UI code, which invited security holes and broke across frameworks. A2UI v0.9 treats the agent as a guest: it describes intent, and the application decides how that intent looks on React, Flutter, Angular, or Lit.

What actually changed in v0.9?

CopilotKit, which collaborated on the spec, says v0.9 is not incremental. The JSON structure, the schema, and the protocol all changed. Communication is now bidirectional, meaning the client can push state back to the agent mid-session.

Google renamed the optional component set from "Standard" to "Basic." The signal is clear: frontend teams should wire agents into the components they already maintain, not adopt a generic kit. A shared web-core library now sits beneath browser renderers, and official React support ships alongside updated Flutter, Lit, and Angular renderers. Community renderers get a dedicated home.

On the agent side, a new Python SDK handles the plumbing. Install it with a single command:

pip install a2ui-agent-sdk

The SDK manages version negotiation, dynamic catalogs that swap schemas at runtime, and resilient streaming. That last feature parses partial LLM output on the fly, healing malformed JSON so components appear as the model generates them.

Why does framework-agnostic matter?

AI agents increasingly need to present information, gather input, and confirm actions. Without a shared contract, every integration becomes bespoke. A Slack bot that shows a date picker in one format breaks when ported to a mobile app expecting SwiftUI.

A2UI positions itself as that contract. Agents speak a portable language; frontends translate it. The same agent can drive a React dashboard, a Flutter mobile app, and a desktop Electron wrapper without rewriting UI logic.

Transports were streamlined too. A2UI now runs over MCP, WebSockets, REST, AG-UI, and the newly launched A2A 1.0. Teams pick the channel that fits their stack.

The skeptics have a point

Not everyone is sold. A Hacker News commenter asked bluntly: "Why the hell would anyone want this? Why on earth would you trust an LLM to output a UI? You're just asking for security bugs, UI impersonation attacks, terrible usability, and more."

The concern is valid. Models hallucinate. Letting one assemble a UI from scratch is a recipe for phishing screens or broken flows. A2UI's answer is the catalog model: the agent can only request components the app explicitly exposes. If an agent asks for something undefined, the request fails.

A Reddit thread pushed back on the opposite front, arguing that fixed catalogs are too restrictive. "Every UI will become the same," one poster warned. Architect Brian Love proposed a middle path: fixed catalogs with dynamic overlays and deterministic fallback when validation fails.

Where A2UI sits in the landscape

A2UI is not alone. Vercel's json-renderer takes a similar declarative approach. AG-UI, MCP Apps, and Oracle's Agent Spec compete for the same niche. Web-focused projects like syntux argue that chatbot-style generative UI produces "disposable" interfaces that lack the consistency and cacheability production apps require.

Google is betting that a framework-agnostic spec can unify the space. The v1.0 release candidate is already published, along with a migration guide from v0.9.1.

Migration and what comes next

Teams on v0.8 face a meaningful upgrade. The evolution guide documents a move to a "prompt-first" schema embedded directly in the model prompt, a refactor into modular files, and a property rename quick reference. The current stable release is v0.9.1, a patch in the v0.9 family.

CopilotKit's Atai Barkai called v0.9 "a BIG improvement, with AG-UI support out the box" and described it as mass-adoption-ready. Whether that adoption materializes depends on how quickly frontend teams see value in wiring their design systems to agents.

Frequently Asked Questions

What is Google A2UI?

A2UI is an open, framework-agnostic specification that lets AI agents declare user interface intent. The host application renders the UI natively using its own design system, avoiding arbitrary code execution.

Which frameworks does A2UI v0.9 support?

Official renderers ship for React, Flutter, Lit, and Angular. A shared web-core library underpins browser renderers, and community renderers are supported.

How do I add A2UI to a Python agent?

Install the A2UI Agent SDK from PyPI with pip install a2ui-agent-sdk. The SDK handles version negotiation, streaming, and partial output healing.

Is A2UI secure?

A2UI uses a catalog model where agents can only request components the app explicitly exposes. This prevents agents from generating arbitrary UI code that could introduce security vulnerabilities.

What transports does A2UI support?

A2UI runs over MCP, WebSockets, REST, AG-UI, and the newly launched A2A 1.0 protocol.

Source: InfoQ