A2UI 1.0: agents describe the screen, your app draws it
Google's A2UI hit 1.0. It is an open protocol for agents that answer with an interface instead of a wall of text: the agent describes what should appear, and each app draws it with its own registered components. The post is dated Nov 11 and was picked up on Hacker News just after midnight UTC on Nov 12: https://developers.googleblog.com/a2ui-10-build-delightful-fast-reliable-agentic-experiences/ (site: https://a2ui.org).
Version 1.0 adds two-way function calls, so a date picker can check availability live instead of waiting for the next model turn, plus accessibility support and an extension system. Renderers exist for React, Angular, Lit, Flutter, SwiftUI and Jetpack Compose, though the two mobile ones are still on 0.9.1 with 1.0 "coming soon." There are agent SDKs for Python, TypeScript and Dart. Data can be bound on the server side, so sensitive values never pass through the model at all.
The quiet big deal: the web and Flutter renderers work inside MCP Apps. One A2UI agent can show the same screen in Claude and in Gemini. Google names real users. Atlassian's Jira agent asks follow-up questions through interactive cards. Alation renders the same data feeds in Claude and in its own product. Bloomberg is building on it. ByteDance has run an A2UI renderer in its Lynx framework since June. Gemini Enterprise lets companies register A2UI agents, though it still supports 0.9.1, not 1.0.
The protocol fits two trends from this month. Agents are getting into other people's apps (Adobe's MCP server, Claude's interactive connectors, Businesses in ChatGPT), and they need a way to show something better than markdown when they get there. And more of the work is choosing from fixed options rather than writing free text: A2UI lets a fast decision model pick components while the app handles the drawing. The open question is whether the major chat apps treat A2UI as a first-class citizen or just one more thing an MCP App can render.
← Back to all articles
Version 1.0 adds two-way function calls, so a date picker can check availability live instead of waiting for the next model turn, plus accessibility support and an extension system. Renderers exist for React, Angular, Lit, Flutter, SwiftUI and Jetpack Compose, though the two mobile ones are still on 0.9.1 with 1.0 "coming soon." There are agent SDKs for Python, TypeScript and Dart. Data can be bound on the server side, so sensitive values never pass through the model at all.
The quiet big deal: the web and Flutter renderers work inside MCP Apps. One A2UI agent can show the same screen in Claude and in Gemini. Google names real users. Atlassian's Jira agent asks follow-up questions through interactive cards. Alation renders the same data feeds in Claude and in its own product. Bloomberg is building on it. ByteDance has run an A2UI renderer in its Lynx framework since June. Gemini Enterprise lets companies register A2UI agents, though it still supports 0.9.1, not 1.0.
The protocol fits two trends from this month. Agents are getting into other people's apps (Adobe's MCP server, Claude's interactive connectors, Businesses in ChatGPT), and they need a way to show something better than markdown when they get there. And more of the work is choosing from fixed options rather than writing free text: A2UI lets a fast decision model pick components while the app handles the drawing. The open question is whether the major chat apps treat A2UI as a first-class citizen or just one more thing an MCP App can render.
Comments