Top AI Product

If you’ve been building anything with AI agents lately, you’ve probably hit the same wall I have: your agent can do incredible stuff behind the scenes, but the moment it needs to present something interactive to the user, you’re stuck with plain text or hacking together custom UI code. Google just dropped something that might actually fix this.

[A2UI (Agent-to-User Interface)](https://developers.googleblog.com/introducing-a2ui-an-open-project-for-agent-driven-interfaces/) went public on February 13, 2026, and it’s been lighting up [The New Stack](https://thenewstack.io/agent-ui-standards-multiply-mcp-apps-and-googles-a2ui/) and [Hacker News](https://news.ycombinator.com/item?id=46286407) since. The idea is deceptively simple: give agents a declarative JSON format to describe UIs, and let the client render those descriptions using its own native widgets. No arbitrary code execution, no iframes, no embedded web views. Just a flat list of pre-approved components — cards, buttons, text fields, date pickers — that the agent references by type and ID.

What makes this actually practical is how LLM-friendly the format is. Because it’s a flat JSON structure with ID references instead of deeply nested trees, models can generate it incrementally. That means progressive rendering works out of the box — your UI builds up piece by piece as the agent streams its response. Think about a travel booking agent that spins up a date picker, a destination selector, and a confirmation button in real time, all rendered as native components on whatever platform you’re running.

The security angle is worth calling out too. A2UI is purely declarative — your app maintains a catalog of trusted component types, and the agent can only request components from that catalog. There’s no path from agent output to script execution, which is a huge deal when you’re letting an LLM drive your UI.

The [GitHub repo](https://github.com/google/A2UI) already ships client libraries for Flutter, Web Components, and Angular, with more renderers on the way from the community. It’s sitting at v0.8 Public Preview under the Apache 2.0 license, so it’s early but very much usable. The [official docs](https://a2ui.org/) have a solid quickstart if you want to get your hands dirty.

People are already calling this the “MCP for agent UIs,” and honestly, the comparison tracks. MCP gave us a standard for how agents talk to tools; A2UI could do the same for how agents talk to users. With Google backing it and [CopilotKit already building integrations](https://www.copilotkit.ai/blog/build-with-googles-new-a2ui-spec-agent-user-interfaces-with-a2ui-ag-ui), this feels like one of those protocols that might actually stick. Whether it wins out over the competing MCP Apps approach favored by Anthropic and OpenAI remains to be seen, but the native-first, no-code-execution philosophy is compelling. Worth keeping an eye on.