iSearch

What is A2A?

A2A, Agent2Agent, is an open protocol that describes how AI agents from different suppliers find each other, communicate and hand over tasks.

Without a common standard, every pair of agents in different systems needs its own connection. A2A offers a shared way to describe what an agent can do and to send it a task.

The protocol was launched in April 2025 and handed over to the Linux Foundation in June that year, so that it is governed independently of any single supplier. The first stable version arrived in March 2026.

Each agent publishes an agent card stating what it can do, where it can be reached and how to sign in. A2A complements MCP, which connects an agent to tools and data.

How it works

An agent that wants help starts by reading the other agent's agent card. It states which tasks the agent can take on, which address it has and which way of signing in it requires. The first agent then sends a task and gets messages and results back.

The task has a course of its own. It can be completed at once, take a long time or pause to wait for more information. Whoever sent the task can follow its progress or be told when the work is done.

Why it matters

The protocol makes it possible for specialised agents to work together, for example an agent in the case management system and one that knows the organisation's procedures. At the same time, every agent gets more parties that can ask it questions, and the one asking is not always a person. That puts the question of whom an AI agent acts for at the centre.

What to keep in mind

The protocol describes how agents sign in but does not decide what they may see. That is settled by access control in the system behind it, which should start from the person the agent works for.

Treat answers from other agents as outside text. They can contain instructions that try to steer the recipient, just as in prompt injection. Log every task in an audit log so that it is possible to follow what happened.

What it means for you

Ask which other agents an AI tool can communicate with and on whose behalf. An agent that may ask questions needs the same permission check as a member of staff.

Common questions

Is A2A the same as MCP?

No. MCP connects an agent to tools and data, and A2A lets agents communicate with each other. Many systems use both. Read more about MCP.

Do we need A2A to use AI agents?

No. The protocol becomes relevant when agents in different systems need to work together. Start with what the agents may do and who approves it. Read more about least privilege.

Sources: The A2A specificationThe Linux Foundation takes over A2A (2025)

All terms Updated