A2A — Agents Talking to Agents
The Agent2Agent protocol: discovery via agent cards, task lifecycles, and where it fits next to MCP.
A2A (Agent2Agent) standardizes how autonomous agents from different vendors and frameworks work together. Announced by Google in 2025 with broad industry backing and now stewarded under the Linux Foundation, it covers discovery — each agent publishes an agent card describing its skills and endpoint — plus task lifecycle (create, progress updates, artifacts, completion) over HTTP/JSON-RPC with streaming. Your travel agent hands the flight subtask to an airline's agent without either side reading the other's code.
The clean mental model against MCP: MCP connects an agent to its TOOLS (vertical — capabilities it wields); A2A connects agents to OTHER AGENTS (horizontal — peers it delegates to). A tool is invoked and returns; an agent negotiates, runs long tasks, streams progress, and can push back. They compose: an orchestrator uses MCP for its own hands and A2A to delegate subtasks to specialist peers, each of which uses MCP internally.
Honest 2026 status: A2A adoption is real but young — strongest in enterprise integration, where cross-org agent handoffs (procurement talking to vendors' agents) justify a protocol. Inside a single codebase you don't need A2A; in-process handoffs are simpler. Reach for it at organizational boundaries — different teams, vendors, trust domains — the same places you'd have reached for a public API contract. And peer agents are exactly as untrusted as any external service: authenticate, authorize, and validate what comes back.