What Is the Agent2Agent Protocol (A2A)?

The Agent2Agent Protocol (A2A) is an open standard for communication between AI agents. It lets an agent find another agent, learn what it can do, hand it a task, and receive progress updates and results, even when the two agents were built by different companies on different frameworks. Each agent stays opaque: it shares its capabilities and its outputs, not its prompts, tools or internal state.

Origin and Governance

Google created A2A, and in June 2025 the Linux Foundation launched the Agent2Agent project to give it vendor-neutral governance (Linux Foundation). In April 2026 the Linux Foundation reported more than 150 supporting organisations and described version 1.0 as the protocol’s first stable specification, adding signed Agent Cards and multi-tenancy (Linux Foundation, April 2026). The project’s site now presents A2A as part of the Agentic AI Foundation (AAIF), the Linux Foundation body that also hosts the Model Context Protocol (a2a-protocol.org). Checked in November 2026.

How A2A Works

The specification defines a small set of concepts:

  • Agent Card. A JSON document that describes an agent’s identity, skills, endpoint and authentication requirements. A public agent usually publishes it at a well-known address on its domain, so clients can discover it. Cards can also be listed in private registries or configured directly.
  • Task. A stateful unit of work with an ID and a lifecycle. Tasks can run for a long time and involve several turns, for example when the remote agent needs more input.
  • Message. One turn of communication from the client or the agent.
  • Part. The content inside a message or artifact: text, a file, or structured data.
  • Artifact. The output the agent produces for a task, such as a report or a data file.

Communication runs over HTTPS. A client can wait for a response, open a stream to receive incremental updates, or register a webhook so the remote agent can push notifications about long-running tasks.

A2A vs MCP

The two protocols solve different problems and are designed to be used together. The A2A site puts it simply: MCP connects an agent to tools, APIs and data; A2A connects it to other agents.

MCPA2A
ConnectsAn agent to tools and data sourcesAn agent to another agent
The other side isA tool server with defined functionsAn autonomous agent with its own reasoning
Typical exchangeCall a function, get a resultDelegate a task, follow its progress, receive artifacts

When to Use A2A and When Not

A2A fits when agents belong to different teams, systems or organisations: a procurement agent that asks a supplier’s agent for a quote, or an internal assistant that delegates HR questions to an HR platform’s agent. It gives both sides a shared contract without either needing to know how the other is built.

Inside one application, where all agents share a codebase, a multi-agent system is usually simpler to build with an orchestration framework calling agents directly. Adding a network protocol between components you own adds latency and failure points without much benefit.

Risks to Plan For

  • Trust. A remote agent’s output is untrusted input. It can be wrong or contain instructions meant to redirect your agent.
  • Authentication and authorisation. The Agent Card states how to authenticate; your system still has to decide what each remote agent may ask for and on whose behalf.
  • Data sharing. Delegating a task sends data to another party. Contracts, data protection rules and data residency still apply.
  • Accountability. When a chain of agents produces a wrong result, logs on both sides are needed to find where it went wrong.

In Practice

The question we ask first is whether A2A is needed at all. If the counterpart is a tool or a dataset, MCP is simpler; A2A earns its place when the counterpart is an agent run by another system or company. In those cases our AI agent development team reads the remote agent’s Agent Card, decides which of its skills to call, sends only the data the task needs and validates each response as untrusted input. Both sides log the exchange under the same task ID, so a delegated task that goes wrong can be followed across the boundary.

Back to Glossary

Let's talk about your project

We'll help you harness AI and ML technologies to create innovative software solutions

File requirements: pdf, doc, docx, rtf, ppt, pptx — max 15 MB

No file selected