Agent2Agent protocolA2A

An open standard that lets artificial intelligence (AI) agents built by different companies find each other, hand off tasks, and send back results, so one agent can bring in another to finish part of a job.

A travel assistant might pass the flight search to an airline's agent and the booking to a hotel's agent, then combine their answers, even though different vendors built each one.

A protocol that runs over standard web technology and treats each remote agent as an opaque service that shares requests and results, never its internal memory or tools. Google introduced A2A in April 2025 and donated it to the Linux Foundation in June 2025. Version 1.0, its first stable release, followed in March 2026, and in August 2026 A2A joined the Agentic AI Foundation (AAIF), the Linux Foundation fund that also hosts the Model Context Protocol.

A typical A2A exchange runs in three steps:

1. Discovery: the calling agent fetches the remote agent's Agent Card, a published file listing its skills, address, and accepted sign-in methods.
2. Authentication: the caller gets a credential, such as an OAuth token, outside A2A and sends it with each request, which the remote agent must check.
3. Task: the caller sends a message, and the remote agent returns status updates and results.

Version 1.0 added Signed Agent Cards, which let a caller check cryptographically who published a card before trusting it.

A2A sits next to terms that are easy to mix up:

• Model Context Protocol (MCP) connects one agent to its tools and data, while A2A connects separate agents across company boundaries. Many systems use both.
• Agent authentication is the broader practice of confirming an agent's identity, and A2A defines one place where it happens.
• Agent Payments Protocol (AP2) extends A2A for purchases, adding signed proof of what the person approved.