Agent2Agent
The Model Context Protocol settled one half of interoperability: how an agent reaches a tool. Agent2Agent — usually written A2A — takes on the other half: how two agents, built by different teams on different frameworks, hand work to each other without either being rewritten as a tool of the other.
How it works
A2A treats a remote agent as a peer that accepts tasks, not as a function to be called. The moving parts:
- An Agent Card — a published JSON document listing the agent's skills, endpoint and authentication, so a client can discover what it is capable of before asking for anything.
- A task as the unit of collaboration, with a lifecycle: submitted, working, input-required, completed or failed.
- Messages and artifacts for the exchange itself, carried over JSON-RPC on HTTP, with streaming updates for work that runs long.
Neither side exposes its internal prompts, memory or reasoning. The remote agent stays opaque — a collaborator, not a subroutine.
Why it needed its own standard
Wrapping another agent as an MCP tool forces a synchronous request-and-reply shape onto something that may run for an hour and may need to come back mid-way to ask a clarifying question. A2A is built for exactly that: long-running, stateful, multi-turn delegation between parties that do not share a codebase. Announced by Google in 2025 with broad vendor backing and since moved under the Linux Foundation, it is the plumbing for a world where your scheduling agent talks to a supplier's agent — and a shared spelling for the handoff, negotiation and task-routing patterns that multi-agent systems otherwise reinvent in private.
Agent2Agent (A2A) is an open protocol for communication between independent agents, letting one delegate a task to another across organisational and vendor boundaries. It defines discovery through the Agent Card, task submission, status updates, and streamed results over HTTP. A2A standardises agent-to-agent delegation the way MCP standardises tool access, so collaborating agents need no bespoke integration built for each peer.
Two companies each have their own software helper. Yours manages your diary; theirs manages their bookings. For your helper to arrange something with them, the two need a shared way of talking — otherwise a person has to sit in the middle copying details across.
Agent2Agent is an open set of rules for exactly that. Each helper publishes a short card saying what it can do and how to reach it. One can then hand a job to the other, receive progress updates, and be asked a question part-way through if something is unclear. Neither has to reveal how it works inside.