A2UI v0.9公開: フレームワーク非依存な生成UIの新標準 A2UI v0.9: The New Standard for Portable, Framework-Agnostic Generative UI

Google has published A2UI v0.9, a new specification aimed at standardizing how AI agents describe and deliver generative user interfaces. As LLM-driven applications increasingly produce UI on the fly rather than emitting plain text, the absence of a common protocol has become a real friction point — and A2UI is positioned as a neutral, portable layer to address it.

At its core, A2UI lets an agent declaratively describe UI components, their state, and the events they emit, leaving the actual rendering to a client that maps those descriptions onto native widgets. This separation means the same UI definition can in principle be rendered on the web, on mobile, or in a desktop shell, without the agent needing to know which framework will ultimately draw it. Version 0.9 reportedly tightens the component model, data binding semantics, and bidirectional event flow, suggesting the spec is approaching production readiness ahead of a future 1.0 milestone.

The broader context matters here. Generative UI has become a competitive design space: Vercel's AI SDK leans on React Server Components to stream UI fragments, Anthropic's Model Context Protocol standardizes how tools and context are exposed to models, and OpenAI's structured outputs are often used to drive bespoke UI renderers. These approaches all work, but each tends to bind the UI representation to a particular stack. A2UI's pitch is that the description format itself should be framework-agnostic, so agents, models, and renderers can evolve independently.

It is also worth reading A2UI alongside Google's earlier Agent-to-Agent (A2A) protocol, which targets interoperability between autonomous agents. If A2A defines how agents talk to each other, A2UI plausibly fills the complementary role of defining how agents talk to humans — through structured, portable interfaces rather than ad hoc HTML or vendor-specific component trees. In a future where multiple agents from different vendors collaborate on a task, a shared UI grammar could be as important as a shared messaging format for delivering a coherent user experience.

Practical questions remain. Adoption will depend on whether major frontend ecosystems ship reference renderers, how gracefully the spec degrades on constrained surfaces such as voice or wearables, and whether security concerns around agent-driven UI — for instance, prompt injection causing misleading interface elements — are adequately addressed. Accessibility is another area where a portable spec must do real work, since semantics that are easy to express in HTML can be harder to preserve when the same description is rendered into a native iOS or Android view.

For developers experimenting with agentic applications today, A2UI v0.9 is likely worth tracking even if it is not yet a default choice. If it gains traction alongside protocols like MCP and A2A, the next generation of AI products may look less like monolithic chat apps and more like loosely coupled agents exchanging both messages and interfaces over open standards. Whether A2UI becomes that standard, or one of several competing options, will become clearer as the spec moves toward 1.0 and as community feedback shapes the remaining design decisions.