operate-a2a-agent-integrationlisted

# Operate A2A Agent Integration Use A2A when independently operated agents need to discover and delegate to each other across process, machine, or framework boundaries. Do not use A2A for editor-to-agent hosting (ACP), agent-to-tool calls (MCP), or in-process subagents. ## Connection Workflow 1. Name the client agent and remote A2A server, including who owns each model, credential set, tool policy, memory, and task store. 2. Fetch the remote Agent Card and verify its advertised endpoint, protocol version and binding, skills, capabilities, and security schemes before sending work. 3. Establish authentication out of band. Treat the card as discovery metadata, not proof that the advertised operator or endpoint is trustworthy. 4. Choose a stateless `Message` for an immediate exchange or a stateful `Task` for long-running, interruptible, or resumable work. 5. Preserve `contextId` for related turns and `taskId` for one task lifecycle. Do not substitute a local conversation or process identifier. 6. Use streaming only when both peers advertise it. Use push notifications only with an authenticated callback, SSRF controls, signature verification, and retry/idempotency rules. 7. Handle `input-required` and `auth-required` as interrupted states that need an explicit follow-up. Treat `completed`, `canceled`, `rejected`, and `failed` as terminal. 8. Validate discovery, authentication, one harmless message, one stateful task, cancellation, streaming or p