# What is Agent to Agent Protocol?

Also called A2A, Agent2Agent.

An agent to agent protocol is a standard that lets independent AI agents, built by different teams on different frameworks, discover each other and exchange tasks over a network. The best known example is A2A, originally published by Google and later donated to the Linux Foundation. Agents describe themselves in a published metadata document and communicate without exposing their internal tools or reasoning.

In A2A, one agent acts as a client and another as a remote agent. The client fetches the remote agent's published description, then sends a task containing one or more messages. The remote agent works on the task, reports a status, and can stream updates as it progresses. Results come back as artifacts, which may include text, files, or structured data.

The central design assumption is opacity. Each agent keeps its own memory, tools, and model choices private, and the protocol carries only tasks, messages, and results across the boundary. That lets an organization expose an agent as a service without revealing how it works internally, which matches how companies already integrate across a trust boundary.

A frequent question is whether such protocols compete with the Model Context Protocol. They address different layers. One connects a single application to the tools and data it drives directly, and the other lets separate autonomous systems delegate work to each other as peers. Many architectures use both at once, with agents exchanging tasks while each uses tool servers privately.

Adoption is early and the specifications are still moving, so any interoperability claim should be checked against a specific version. Delegation across agents also introduces failure modes that single agent systems do not have, including tasks that stall inside a remote system, ambiguous ownership of retries, and cost that compounds as agents call other agents.

## Key points

- Lets agents from different vendors delegate tasks to each other.
- A2A was published by Google and donated to the Linux Foundation.
- Agents stay opaque; internal tools and reasoning are not shared.
- Complements rather than replaces tool connection protocols.
- Specifications are young, so check interoperability against a version.

## In practice

A travel planning agent needs visa rules it does not cover. It fetches the published description of a compliance agent at another company, sees a skill for checking entry requirements, and sends a task naming the traveler's passport country and destination. The compliance agent streams progress, then returns a written summary as an artifact. Neither agent learns which model or internal tools the other used.

## Related terms

- [Agent Card](/en/glossary/agent-card)
- [Model Context Protocol](/en/glossary/model-context-protocol)
- [Webhook](/en/glossary/webhook)
- [REST API](/en/glossary/rest-api)
- [API](/en/glossary/api)

[Back to the AI Glossary](/en/glossary)