A2A

Editorial note

The draft has improved meaningfully in the past two quarters. The political question — does a Google-led standard land in this layer? — is still open.

Our analysis

A2A's premise is that agents will need a shared way to discover, authenticate, and exchange tasks with each other in a way that doesn't reduce to "call an MCP server." The proposal models an agent as an addressable entity with a capability manifest, a task lifecycle, and a delegation primitive. The model is right in shape; the open questions are at the same seams that MCP is fighting at — identity, audit, and the boundary between protocol-level streaming and application-level state. The draft has improved meaningfully across versions 0.2 to 0.4: the capability manifest format is now stable enough to write client libraries against, and the task-cancellation semantics are no longer underspecified. What remains contested is whether the standard belongs in a Google-stewarded working group or whether a more neutral body (a working group hosted under one of the foundation umbrellas) is going to fork it. The choice will be made by adoption, not by working-group politics. We think A2A reaches Active by mid-2027 if Anthropic and OpenAI commit; otherwise it stalls at Draft 0.6.

What we are watching

• Foundation-host conversation (CNCF? LF AI?)
• Anthropic and OpenAI reference implementations
• Whether the capability manifest format absorbs OpenAgents work

How to cite this entry

Citation format: The Agentic Review. "A2A — Agent-to-Agent Protocol." Standards Tracker, last updated 2026-05-02. https://agentic.review/standards/a2a/. See our citation style for additional guidance.

Update history

Material updates to this entry are dated and logged in our corrections log. The version above reflects the most recent review by the standards desk.