A2A stable

What it is

A2A lets one agent delegate work to another agent it does not control, across teams or companies, without either side exposing its internals. Google announced it in April 2025, donated it to the Linux Foundation in June 2025, and v1.0 shipped in March 2026.

How it works

• Each agent publishes an Agent Card, a JSON document at a well-known URL describing its skills, endpoints and auth requirements. v1.0 added signed Agent Cards so a card can be verified across organizational boundaries.
• A client agent opens a task. Tasks have a lifecycle (submitted, working, input-required, completed, failed) and can run for hours.
• Messages carry parts (text, files, structured data). Results come back as artifacts.
• Transport is JSON-RPC over HTTP, with SSE for streaming and push notifications for long jobs.

A2A is available natively in Microsoft Copilot Studio, Azure AI Foundry and Amazon Bedrock AgentCore, and more than 150 organizations back it.

When to use it

Your agent needs something done by an agent owned by someone else, and you want a contract rather than a shared codebase. If the other side is a tool with a fixed API, use MCP instead. The two are complementary: an A2A agent typically uses MCP tools internally.

What changed

• March 2026: v1.0, the first stable release. Signed Agent Cards.
• June 2025: moved to the Linux Foundation.

Watch out for

• Discovery is not solved. Agent Cards tell you how to call an agent, not which agent to trust.
• Long-running tasks mean you need durable state on your side too.