Agent2Agent (A2A)

An open protocol that lets independently operated agent systems discover one another, exchange messages and track work as a task with its own state and output artefacts.

Updated 15 September 2026

In practice

A research agent hands a bounded extraction task to a specialist agent run by another company, disconnects, reconnects an hour later and asks for the task's current state.

The distinction

A2A moves tasks between agents; MCP connects one application to its tools. A2A carries the conversation and task state, and leaves the principal, permission, quality, contract and remedy to the systems around it.

On this page

The task is the centre

A2A puts a Task at the middle of the exchange: a server-generated identifier, a current status, artefacts, history and metadata. The caller can submit work, disconnect, reconnect and ask for the state.

That solves a genuine transport problem. A remote agent can request more input, report progress or return a file without exposing its own prompt, memory or tool implementation.

Completed is not accepted

The protocol's completed state means the remote implementation reached its terminal success state. It does not mean the parent agent accepted the file, the customer received what was promised, or a downstream record changed correctly.

That gap is the one to design for. A child task can complete while the parent discovers the schema is wrong, the evidence is stale or the customer needed something else. Acceptance is a separate record, owned by whoever is entitled to make it.

What sits beside the wire

Five questions stay outside the message: which person or organisation owns the objective; what this actor may do with the data and budget; what makes the returned work correct; whether this is a one-off request or a supplier contract; and who repairs a loss.

A signed message can show integrity. It cannot show that the sender had the right to commission the work.

Sources & context

A2A Protocol Specification

A2A Project · 1.0

Defines Agent Cards, messages, tasks, artefacts, streaming, push notifications and task operations, and leaves the authorisation model to the agent and its operator.

Source checked 2026-09-13

Announcing the Agent2Agent Protocol

Google Developers Blog · 2025

States the design intent that independent agents collaborate without sharing internal state, memory or tools. A launch announcement, not adoption evidence.

Source checked 2026-09-13