Sources: Shubham Saboo's LinkedIn post announcing the 1.0 release (2026-11-11); A2UI v1.0 specification; earlier Google Developers Blog posts on the launch and v0.9.
In one paragraph
A2UI is an open (Apache 2.0) protocol, started by Google, that lets an agent reply with interactive interface instead of text: forms, charts, pickers, approval cards. The agent sends a stream of declarative JSON describing what to show, and the host app renders it with its own native components from a catalog the app controls. The agent can't ship code or use components it wasn't given, so screens match the product and the attack surface stays small. Version 1.0 is the first stable release. It ships renderers for React, Angular, Flutter and SwiftUI, agent SDKs for Python, TypeScript and Dart, and built-in accessibility; native Android is still to come. Announced adopters include Atlassian (decision cards inside Jira work items), Alation, ByteDance's Lynx framework, Bloomberg, and Google's own Gemini Notebook and Gemini Enterprise. The positioning: MCP connects agents to tools, A2A connects agents to each other, and A2UI covers what the user actually touches.
How the protocol works
Surfaces and a message stream. The agent-to-renderer channel has four message types: createSurface, updateComponents (add or update component definitions), updateDataModel (insert or replace a surface's data), and deleteSurface. The renderer builds the UI progressively as messages arrive, so a screen appears while the agent is still working. The transport layer marks the end of an agent turn.
Structure separate from data. Components bind to a data model through paths, so values can be filled in server-side and sensitive data never has to pass through the model. Surfaces can pull live data while open, such as availability in a booking picker.
Client-owned catalogs. The renderer offers a catalog of components. In 1.0, components can mix catalogs per component (@catalogId), and catalogs accept standard JSON Schema metadata.
New in 1.0. Wire envelopes are renamed agent_to_renderer and renderer_to_agent. Bindings and system keys move to an @ namespace (@path, @call, @accessibility, @metadata, @checks). There are typed, bidirectional function calls checked against declared signatures (callRendererFunction, callAgentFunction and their responses), single-message createSurface, strict Unicode identifier rules, and extension points for security metadata, access policies and telemetry at surface and component level.
Where it sits among agent UI standards
A2UI describes what to render, declaratively, from a trusted catalog. Event protocols such as AG-UI (R034) describe the run itself: streamed text, tool calls, state deltas and interrupts. MCP Apps embed tool-provided HTML. The approaches can be combined, for example A2UI surfaces carried over an event stream. They differ mainly in who owns the rendering code and how much the agent can inject.
Tradeoffs and failure modes
Catalog design is now product work. The agent is only as expressive as the components you expose, and catalogs drift between apps and platforms.
Model reliability. The agent must produce valid component trees and bindings. Validation and graceful fallback in the renderer are essential.
Bidirectional calls widen the boundary. Renderer-to-agent and agent-to-renderer function calls need the same authorization thinking as any API.
Breaking changes from 0.9. Renamed envelopes and the @ prefixes mean existing integrations need migration, which the evolution guide covers.
Vendor-led standard. Adoption is broad, but governance and the pace of change are still set mainly by one company.
Review checklist
Which components are in your catalog, and who approves new ones?
Are sensitive values bound server-side so they never enter the model's context?
Does the renderer validate every message against the 1.0 schemas and degrade cleanly to text?
Are callable functions in both directions declared, typed and authorized?
Do surfaces meet accessibility requirements, using the protocol's accessibility fields?
How does an approval card map to a durable, auditable decision in your system?
Related
R034 (AG-UI): the event-stream protocol between agents and user-facing apps, a complementary layer.
R016 (do humans still need to check AI's facts?): approval and verification moments are a main use for interactive cards.
R006 (six harness terms): the harness edge where agent output reaches the user.