# A2A protocol

The A2A protocol (Agent2Agent) is an open standard that lets AI agents built by different teams or vendors discover each other, hand off tasks and exchange results, without exposing their internal logic, memory or tools.

## What is the agent to agent protocol used for?

The agent to agent protocol solves a coordination problem. Companies now run many specialized agents (one for billing, one for IT requests, one for research), built on different frameworks by different teams. Without a shared protocol, each pair needs custom integration code.

A2A gives them one way to talk. An agent can find another agent, see what it can do, send it a task and get the result back, without knowing how that agent works inside. Google introduced A2A in April 2025, and it is now an open source project under the Linux Foundation.

## How does the A2A protocol work?

The [specification](https://a2a-protocol.org) defines a client agent, which sends a request on behalf of a user, and a remote agent, which does the work. Four concepts carry the exchange:

- **Agent card:** a JSON document the remote agent publishes, listing its name, endpoint, skills and the authentication it requires. Client agents read it to decide whether to delegate.
- **Message:** one turn of communication, from the user side or the agent side, made of parts (text, a file or structured data).
- **Task:** a unit of work with an ID and a lifecycle. It can be working, need more input, complete or fail, which suits jobs that take minutes rather than milliseconds.
- **Artifact:** the output of a task, such as a report or a file.

Requests can use any of the three bindings in version 1.0 of the spec: JSON-RPC 2.0, gRPC or HTTP+JSON. Long tasks can stream updates over server-sent events or send push notifications to a webhook. The remote agent stays opaque: the client sees its skills and results, never its prompts, memory or tools.

Example: a user in an HR platform asks, "Is my laptop request approved yet?" The HR agent doesn't own IT tickets, so it reads the IT agent's card, sends it a task with the employee ID, and receives a completed task with the ticket status. The user gets one answer in one place.

## What is the difference between A2A and MCP?

They solve different problems and work together.

| | MCP | A2A |
|---|---|---|
| Connects | An AI application to tools and data | One agent to another agent |
| The other side is | A server exposing functions and resources | An independent agent with its own reasoning |
| Typical call | Run this query, read this file | Handle this task and report back |
| Introduced by | Anthropic (Nov 2024) | Google (Apr 2025) |
| Governed by | Linux Foundation (Agentic AI Foundation) | Linux Foundation |

A common design uses both: an agent reaches its own tools over MCP and delegates whole tasks to other agents over A2A.

## How do in-product agents use A2A?

An agent inside a SaaS product can't own every job a user brings. A2A lets it pass the parts that belong to another system and keep the user in one conversation. In Foldspace, the Product Agent does this only when a request calls for one of the customer's own agents: it [hands that agent the task over A2A and returns the result](https://docs.foldspace.ai/user-guides/a2a-agents/) to the same conversation.

## Learn more

- [A2A quickstart](https://docs.foldspace.ai/guides/a2a-quickstart)