How do agents talk to each other (A2A)?

You book a flight, and before you pay, the airline's site offers a hotel and a rental car for the same dates. Those come from separate companies. Behind the page, their systems pass your dates and city back and forth in a format they agreed on ahead of time. Artificial intelligence (AI) agents now need the same kind of agreement. One of the best known is called Agent2Agent (A2A).

Topic 89 described a team of agents inside one system, all built by the same company. Many real errands cross company lines. Your assistant might come from one firm, while a hotel's booking agent, or a supplier's ordering agent, comes from another. They were built with different tools, by people who never met. For them to work together, something has to set the ground rules.

Without shared rules, every pair of agents needs its own custom connection. Ten agents that all want to talk to each other would need forty-five separate links, each one built and kept up by someone. A shared protocol, a set of rules everyone follows, changes the math. Each agent learns the rules once, and can then work with any other agent that follows them too. It's the same reason your phone can call any other phone, whoever made it.

A2A is an open protocol that lets an agent from one company hand a task to an agent from another, and get the result back. Google announced it in April 2025, with more than fifty partner companies. That June it moved to the Linux Foundation, a nonprofit that looks after shared software, so no single company runs it. By April 2026 the project reported more than 150 supporting organizations.²⁴⁵

Before one agent can hand work to another, it has to find it and know what it can do. In A2A, an agent can publish an agent card. That's a short public file, a bit like a business card, listing the agent's name, the skills it offers, where to reach it, and how a caller should prove who it is. Another agent reads the card and decides whether this is the right helper for the job.³

Once they've found each other, the asking agent sends a task. The other agent takes it, works on it, and reports where it stands: submitted, working, or waiting for more information. If it needs something from you, such as which of two dates you prefer, it can pause and ask. When it's done, it sends the result back.³

You may have heard of the Model Context Protocol (MCP) as well, in topic 28. The two do different jobs. MCP connects an agent to tools and data, like a calendar or a company's files. A2A connects agents to each other. Google described A2A as a complement to MCP, so a single agent can use both: MCP to reach its tools, and A2A to call on another agent for help. A2A also had a rival. The Agent Communication Protocol (ACP) did much the same job until it merged into A2A in August 2025.¹²⁶⁷⁸

Shared rules make the hand-off possible. They don't decide what an agent should be allowed to do. That's still up to you and the people who build the tools: the task you set, the limits you give, and the points where a person approves a step before it counts. Topic 60 covers that approval habit. If an assistant offers to pass part of a job to another company's agent, ask which agent, and what it gets to see.