Google released A2UI v0.9 yesterday, and the core shift is worth naming directly: agents no longer invent UI. Instead, they declare intent—button clicks, form fields, data tables—against a catalog of components you already own. That's not a small philosophical change; it's a hard constraint on what an LLM can output, and it cuts boilerplate for teams shipping agentic apps.
The Catalog Model: Agents Speak Your Design Language
Previous versions of A2UI shipped a "Standard" component set — a generic palette of buttons, inputs, modals. v0.9 renames it "Basic" and pivots the design. The message: you should wire your agent to the components you've already built. If you have a custom date-picker, a branded alert, a data-grid with your company's accessibility rules baked in — those are what the agent references. The agent outputs a JSON intent — "show a form with these fields, call this validation, sync this data back" — and your client-side renderer maps that intent to the components in your catalog.
The consequence: no security theater where LLMs dream up UI shapes. No "UI impersonation attacks" or unstyled surprises. The component boundary is enforced at parse time. If the agent tries to emit a component name not in your catalog, validation fails deterministically. CopilotKit's Atai Barkai called v0.9 "mass-adoption-ready" for this reason—it closes the loophole.
Installation and the SDK
To add the agent-side toolkit to a Python agent:
pip install a2ui-agent-sdk
The SDK handles the work you'd otherwise hand-code: version negotiation with the client, dynamic catalog switching (swap schemas at runtime if your frontend updates), and resilient streaming that parses partial LLM output incrementally so components render as they arrive, not after the full response lands.
On the client side, v0.9 now ships an official React renderer—alongside upgraded Flutter, Lit, and Angular renderers. A shared web-core library underpins all browser implementations, reducing drift. There's also a dedicated home for community renderers if you need a Svelte or Vue adapter.
The Bigger Shift: Bidirectional and Modular
v0.9 is not a minor update. The JSON structure changed. The schema changed. The protocol became bidirectional—agents can now request data or validation feedback from the client, not just push UI declarations downward. New language features include:
- Client-defined validation functions — your browser enforces rules the agent declares, not the other way round.
- Client-to-server data syncing for collaborative editing — forms that update in real-time.
- Simplified, modular schema — instead of one monolithic spec, features are now composable.
Transports are also streamlined. A2UI now speaks MCP, WebSockets, REST, AG-UI, and the newly launched A2A 1.0. If you're already running MCP servers for your agents, A2UI can ride over that without new infrastructure.
Migration and What's Shipping Today
Google published a v0.8 to v0.9 evolution guide covering the move to a "prompt-first" schema, the refactor into modular files, and property renames. The current stable release is actually v0.9.1 (a patch). A v1.0 release candidate specification is already published with its own evolution guide, so teams planning multi-year commitments should scan that.
The Skepticism and the Trade-off
Not everyone is convinced. Hacker News pushed back: "Why on earth would you trust an LLM to output a UI?" Reddit raised concerns that the catalog model is too restrictive—"every UI will become the same." Architect Brian Love proposed a pragmatic middle ground: fixed catalogs with dynamic overlays and deterministic fallback.
The trade-off is real. A2UI buys you predictability and security at the cost of flexibility. Your agent can't invent novel layouts. It can't explore the design space. It's constrained to what's in the catalog. For internal tools, admin dashboards, and business agents—that's a feature. For creative applications or specialized UX—it's a ceiling.
A2UI isn't alone. It sits alongside AG-UI, MCP Apps, Vercel's json-renderer, and Oracle's Agent Spec. But Google is positioning A2UI as the portable contract sitting beneath all of them—a lowest common denominator so agents can emit UI portably across frameworks and transports.
If you're shipping agentic products and worried about UI consistency, validation, or security, v0.9 is worth a sprint to evaluate. The SDK is one pip install away.