a2a

# A2A — Agent-to-Agent Protocol ## Overview A2A is an open protocol from Google that defines how AI agents communicate, delegate tasks, and collaborate with each other. While MCP focuses on tool integration, A2A addresses agent-to-agent interoperability. ## Core Concepts ### Agent Card Every A2A agent publishes an **Agent Card** at `/.well-known/agent.json` describing its capabilities: ```json { "name": "Research Agent", "description": "Finds and summarizes information on any topic", "url": "https://research-agent.example.com", "capabilities": { "streaming": true, "pushNotifications": true }, "skills": [ { "id": "web-research", "name": "Web Research", "description": "Search and summarize web content" } ] } ``` ### Task Lifecycle A2A communication revolves around **Tasks** with a defined lifecycle: ``` submitted → working → [input-required] → completed / failed / canceled ``` ### Messages and Parts Agents exchange **Messages** containing **Parts** (text, files, data): ```json { "role": "agent", "parts": [ { "type": "text", "text": "Here is the research summary." }, { "type": "file", "file": { "name": "report.pdf", "mimeType": "application/pdf", "bytes": "..." } } ] } ``` ## Transport A2A uses HTTP + JSON-RPC 2.0: - **`POST /`** — send JSON-RPC requests - **SSE** — streaming responses for long-running tasks - **Push notifications** — webhook callbacks for async completion ## Key Methods | Method | Description | |--------|-------------| | `tasks/send` | Send a message to create or continue a task | | `tasks/sendSubscribe` | Send and subscribe to streaming updates | | `tasks/get` | Get current task state | | `tasks/cancel` | Cancel an in-progress task | | `tasks/pushNotification/set` | Register a webhook for task updates | ## Example Flow ``` Client Agent │ │ │── tasks/send ────────────────►│ Create task with user message │◄── status: working ──────────│ Agent begins processing │◄── status: completed ────────│ Agent returns result │ │ ``` ## Multi-Agent Collaboration A client agent can delegate subtasks to specialist agents: 1. Discover agents via their Agent Cards 2. Send tasks to the most suitable agent 3. Receive results and incorporate into the parent task 4. Handle `input-required` status for interactive collaboration ## A2A vs MCP | Aspect | A2A | MCP | |--------|-----|-----| | Focus | Agent-to-agent communication | Agent-to-tool integration | | Discovery | Agent Cards (`.well-known/agent.json`) | Server capabilities negotiation | | Transport | HTTP + JSON-RPC + SSE | stdio, HTTP + SSE | | Interaction | Task-based (long-running) | Request-response (tool calls) | | Complementary | Delegates to agents | Calls tools | ## Best Practices - Publish an Agent Card with accurate skill descriptions so other agents can discover and route tasks correctly. - Use streaming (`tasks/sendSubscribe`) for long-running tasks to provide progress updates. - Handle the `input-required` status to support interactive multi-turn workflows. - Combine A2A with MCP — use A2A for agent-to-agent delegation and MCP for tool integration within each agent. - Implement idempotency on task IDs so retries don't create duplicate work.