Agent2Agent Protocol (A2A)

A2A gives an AI agent two things it did not previously have in a standard form: a way to announce what it can do, and a way to hand off a task to another agent and get a result back.

An A2A-compliant agent publishes an Agent Card — a JSON document, conventionally served at /.well-known/agent-card.json — that declares its name, the "skills" it offers, the endpoint to reach it at, and the authentication it requires. Another agent (or the client acting on a user's behalf) reads that card before sending it any work, the same way a browser reads a robots.txt or a manifest before acting on a site.

Once discovery is done, the requesting agent opens a task with the remote agent over JSON-RPC 2.0 (gRPC is also supported for lower-latency internal use). A2A supports synchronous request/response for quick tasks, streaming updates over SSE for work that reports partial progress, and asynchronous push notifications for tasks that run long after the requesting agent has moved on. The remote agent does the work — which might itself involve calling tools over MCP — and returns a structured result, an artifact, or a request for more input.

Google announced A2A in April 2025 as a protocol for letting agents built on different frameworks and by different vendors interoperate, rather than requiring every pair of agents to be custom-wired together. In June 2025 Google donated the protocol, its specification, and its SDKs to the Linux Foundation, which set up a dedicated Agent2Agent project to run it under vendor-neutral governance instead of a single company's roadmap.

That governance moved again in August 2026: A2A became a hosted project of the Agentic AI Foundation (AAIF), the Linux Foundation-directed fund that already stewards MCP. The two protocols now sit under the same neutral roof, which is a reasonable signal that the "agent talks to tools" layer and the "agent talks to agent" layer are being treated as one connected stack rather than competing standards. By its first anniversary in April 2026, A2A had more than 150 supporting organizations, and it ships as a native option in Google Cloud, Microsoft's Azure AI Foundry, and AWS's Bedrock AgentCore.

These get conflated constantly, and the distinction is layer, not competition:

Protocol	Connects	Governed by	Commerce-specific?
A2A	One agent to another agent	Agentic AI Foundation (Linux Foundation)	No
MCP	An agent to its tools and data	Agentic AI Foundation (Linux Foundation)	No
ACP	A buyer's agent to a merchant, for checkout	OpenAI and Stripe, foundation path stated as future	Yes
UCP	A buyer's agent to a merchant's full commerce surface	Google-led	Yes

MCP is vertical: model to tools. A2A is horizontal: agent to agent. Neither one moves a product or takes a payment — that is what ACP and UCP are for, and UCP explicitly lists A2A as one of its supported transports for agent-to-agent interaction alongside REST, JSON-RPC, and MCP. A shopping or procurement agent could plausibly use all four in one purchase: MCP to look up your product data, A2A to hand a sub-task to a specialist pricing or logistics agent, and ACP or UCP to actually check out.

A2A does not touch your catalog directly — no agent asks your systems anything over A2A unless you are operating your own agent. But it is part of the same shift that makes MCP and the commerce protocols relevant: buyers are increasingly represented by agents that discover, delegate, and act on their behalf, sometimes chaining a general shopping agent to a specialist procurement or price-comparison agent before a human ever sees a result.

Whichever protocol carries the request — a search crawler, a chat answer engine citing your page, an MCP-connected assistant, or an A2A hand-off between two agents — what gets returned is bounded by what your product record actually contains. A protocol cannot invent a pressure rating, a certification, or a compatible part number that your data doesn't have. The protocol layer is being solved by the industry; the record underneath it is still each seller's own job.