The Agent2Agent (A2A) protocol: getting AI agents to talk to each other

The idea in one sentence

MCP standardised how an agent plugs into tools. A2A covers the other direction: how one AI agent calls on another one, which it neither knows nor controls, to delegate part of the work.

How it works

The official documentation describes a handful of simple building blocks:

• Agent Card: a JSON document each agent publishes, describing its identity, skills, endpoint and required authentication. This is what makes discovery possible.
• Task: a unit of work with an ID and a lifecycle (working, completed, waiting for input…), designed for long-running jobs.
• Messages and artifacts: the exchanges between agents and the deliverables they produce (a document, an image, structured data).

Traffic runs over HTTP(S) with JSON-RPC messages, with streaming for incremental results and push notifications for long tasks. The remote agent stays opaque: it shares neither its memory nor its tools.

A concrete example

An in-house recruiting agent needs to set up interviews. It reads the Agent Card of a scheduling agent from another vendor, sends it the task “find three slots with these candidates”, receives updates while the remote agent works with the calendars, then collects the list of slots as an artifact. Neither side needs to know which framework the other uses.

Origin and governance

Google announced A2A on April 9, 2025 on its developer blog, backed by more than 50 technology partners. On June 23, 2025 the Linux Foundation launched the Agent2Agent project, to which Google contributed the protocol, with AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow among the founding members. The specification has since reached a stable 1.0 version. On August 27, 2026 A2A was accepted as a “Growth Stage” project of the Agentic AI Foundation (AAIF), a Linux Foundation-directed foundation that also hosts MCP, goose and AGENTS.md.

Why it matters, and common myths

For agent orchestration across organisations, a shared protocol saves writing a custom integration for every pair of agents. Two misconceptions come up often:

• “A2A is a Google product.” Google launched it, but governance is now neutral and shared among several companies.
• “With A2A, agents just cooperate on their own.” The protocol is plumbing: the quality of a multi-agent system still depends on task design, the permissions granted and human checks.

To compare frameworks that support these standards, see the agent frameworks page.