What is the A2A protocol?
A2A (Agent2Agent) is an open protocol from Google for letting independent AI agents discover each other, exchange structured messages, and delegate tasks over HTTP. It sits next to MCP in the stack: MCP wires an agent to tools, A2A wires it to other agents.
A2A is Google's open protocol for agents to talk to other agents. It was announced April 9, 2025 at Google Cloud Next with around 50 launch partners including Salesforce, SAP, and Atlassian [unverified], and now sits under the Linux Foundation. The shift it forces: MCP connects agents to tools, A2A connects agents to other agents. Both live in the same stack.
How it works
Each agent exposes an AgentCard at /.well-known/agent-card.json describing its capabilities and endpoint. Clients call those endpoints over HTTP using JSON-RPC 2.0. Payloads carry text, files, or structured JSON, and a session can run synchronous, stream via Server-Sent Events, or push asynchronous updates.
The non-obvious bit
A2A specifies a task state machine for long-running work: submitted, working, input_required, completed, plus terminal states canceled, rejected, failed. A caller agent does not have to assume the remote agent will answer in one round trip. It polls or subscribes while the task moves through states, and can be asked for more input mid-task. That lifecycle is what makes cross-agent delegation work beyond simple request and response.
Reference SDKs ship for Python, Java, Go, JavaScript, and .NET under Apache 2.0.
Last updated: May 20, 2026