
A2A protocol: how AI agents talk to other agents
In short: Agent2Agent (A2A) is an open protocol that lets AI agents from different vendors discover each other and hand off tasks, while MCP connects agents to tools.
The Agent2Agent (A2A) protocol is an open standard for AI agents built by different teams or vendors to discover one another, describe what they can do and exchange tasks and results. Google announced it on 9 April 2025 with more than 50 technology partners and later donated it to the Linux Foundation. MCP connects an agent to tools, and A2A connects it to other agents.
A2A protocol in plain words
MCP is how an agent picks up a tool. A2A is how one agent phones a colleague in another company and says, "can you handle this part?"
How it works
Each agent publishes an Agent Card, a JSON description of its capabilities. A client agent finds the card and sends a task. The remote agent works on it and returns the output, which the protocol calls an artifact. Each task has a lifecycle, so the two agents can stay in sync on long-running work.
The remote agent's internals stay opaque. The A2A documentation, read in November 2026, says agents collaborate without sharing internal memory, tools or proprietary logic. It lists version 1.0 of the protocol and official SDKs in six languages.
Example
As an illustration, a travel company's planning agent sends a task, "hold two rooms near the venue for 12 to 14 March", to a hotel group's booking agent via A2A. The booking agent replies with options and a hold reference. Neither side shares prompts, tools or data stores.
Where you'll meet it
- Enterprise agent platforms that want cross-vendor interoperability
- Partner integrations between companies' agents
- SDKs and samples in the A2A project
- Architecture discussions alongside MCP
What it changes for you
For architects, it is a way to integrate agents without merging codebases or exposing internal tools. For security teams, it creates agent-to-agent trust questions: who authenticated, what was delegated and who is accountable when the remote agent is wrong. My view: treat it as an option to evaluate for each integration, and do not adopt it by default.
A2A vs MCP
With MCP, an agent calls a tool and the client controls the call. With A2A, an agent asks another agent, and the remote side keeps its own reasoning and tools. The A2A documentation's own advice is to use both.
Common mistake
My view: using A2A between agents you own in the same codebase. A function call or a subagent is simpler. A2A earns its overhead across team or company boundaries.
Related terms
Model Context Protocol · Multi-agent system · AI agent · AI governance
Use A2A across team or company boundaries, and a function call or a subagent inside your own codebase.


