What is the A2A (Agent2Agent) protocol?

A2A stands for Agent2Agent, a protocol that lets one AI agent hand work to another. If you run into A2A in a payments document, it may mean something else: account-to-account transfers, which have nothing to do with this. What matters to a store is what sits underneath what it sees: A2A is one of the roads the agentic commerce protocols are built on, which is why it appears in the specifications for UCP (Universal Commerce Protocol) and AP2 (Agent Payments Protocol) even though the shopper never sees it.

How it works

The specification defines it as an open standard for communication between independent AI agent systems, built with different tools, languages or vendors. It has four goals: agents discover what each other can do, agree on how to exchange information (text, files or JSON data), manage tasks together and exchange information securely, without access to the other's internal state, memory or tools.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)} Google explained it with two roles: a client agent formulates and communicates the task, and a remote agent carries it out, either to provide information or to take an action.⁠^{[2]Google Developers Blog (April 9, 2025). Announcing the Agent2Agent Protocol (A2A)}

The calling card. Every A2A server has to offer an Agent Card: a JSON document with its identity, its capabilities, its skills, the addresses it listens on and what it requires to authenticate. A client finds it in one of three ways: at the well-known address https://{domain}/.well-known/agent-card.json, by querying agent catalogs, or because someone configured the address by hand.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

The task. Work is organized into tasks with states: submitted, working, completed, failed, canceled, rejected, and two waiting states, for when the agent needs more input from the person or an additional authorization. Agents send each other messages made of parts (text, file references or JSON), and a task's result arrives as artifacts, which the specification keeps separate from messages.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

How the answers arrive. A task can be answered in the same request, live with Server-Sent Events, or later, with notifications the remote agent sends as a webhook to an address the client registered. It is built for long-running tasks, including ones with a person in the loop.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

What it travels over. Version 1.0 defines three protocol bindings: JSON-RPC 2.0, gRPC and HTTP with JSON. And it supports the usual web security schemes: API keys, HTTP authentication, OAuth 2.0, OpenID Connect and mutual TLS.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

From Google to the Linux Foundation

Google introduced A2A on April 9, 2025, with more than 50 technology companies behind it, including Atlassian, PayPal, Salesforce, SAP and Workday, and described it from the start as a complement to Anthropic's MCP (Model Context Protocol).⁠^{[2]Google Developers Blog (April 9, 2025). Announcing the Agent2Agent Protocol (A2A)} On June 23, 2025, at Open Source Summit North America in Denver, it donated A2A to the Linux Foundation so it could be a neutral, community-governed standard.⁠^{[3]Google Open Source Blog (April 16, 2026). A year of open collaboration: Celebrating the anniversary of A2A}

Version 1.0 came out in March 2026, the first stable, production-ready release, with signed Agent Cards that verify an agent's identity cryptographically.⁠^{[3]Google Open Source Blog (April 16, 2026). A year of open collaboration: Celebrating the anniversary of A2A} The specification published today is 1.0.0, and the earlier ones (0.3.0, 0.2.6 and 0.1.0) remain accessible.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)} At its first anniversary, on April 9, 2026, the Linux Foundation counted more than 150 organizations backing it and placed it inside services from Microsoft, Google and AWS.⁠^{[4]Linux Foundation, press release (April 9, 2026). A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms, and Sees Enterprise Production Use in First Year}

Where it touches commerce

A store does not run into A2A on its own, but inside the shopping protocols. AP2 presents itself as an open extension of A2A: agents collaborate over A2A and use AP2 to secure payments.⁠^{[5]AP2, official protocol site. AP2 - Agent Payments Protocol Documentation} UCP supports four transports, REST, MCP, A2A and one for embedded experiences, and when a merchant declares A2A, the address it publishes is that of its agent's Agent Card.⁠^{[6]Universal Commerce Protocol. UCP Specification Overview (version 2026-08-25)} Google sums up the relationship in its first-year review: AP2, UCP and A2UI, a protocol for agent interfaces, were all designed with A2A's extension model.⁠^{[3]Google Open Source Blog (April 16, 2026). A year of open collaboration: Celebrating the anniversary of A2A}

In a store

Here is how the agent of a store that answers catalog questions over A2A might introduce itself. It is a fictitious example, written with the fields of the 1.0 specification; a real card also declares its security schemes:⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

{
  "name": "Northpack Backpacks catalog agent",
  "description": "Answers questions about the catalog: dimensions, materials, stock and shipping times.",
  "supportedInterfaces": [
    {
      "url": "https://northpack.example/a2a/v1",
      "protocolBinding": "JSONRPC",
      "protocolVersion": "1.0"
    }
  ],
  "provider": {
    "organization": "Northpack Backpacks (fictitious example)",
    "url": "https://northpack.example"
  },
  "version": "1.0.0",
  "capabilities": { "streaming": true, "pushNotifications": false },
  "defaultInputModes": ["text/plain"],
  "defaultOutputModes": ["text/plain", "application/json"],
  "skills": [
    {
      "id": "search-products",
      "name": "Search products",
      "description": "Finds backpacks by capacity, use and maximum price.",
      "tags": ["catalog", "search"],
      "examples": ["Waterproof 20-liter backpack for a 15-inch laptop under $90"]
    }
  ]
}

A person's shopping agent would read this card at /.well-known/agent-card.json, see that there is a search skill, send a message with the request to the declared address and get back a task that moves to "working" and ends with an artifact: the list of backpacks, as text or JSON. At no point does it see how the store's agent queries the catalog.

How to check it

1. See whether your platform or your payment gateway declares A2A. In a UCP profile, the a2a transport points to the URL of an Agent Card.⁠^{[6]Universal Commerce Protocol. UCP Specification Overview (version 2026-08-25)}
2. Request the card. On a domain that says it has an A2A agent, https://{domain}/.well-known/agent-card.json has to return an AgentCard object.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}
3. Read the skills, the input and output modes and the security. They are what the agent promises to do and what it requires in order to do it.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}
4. Check the version in protocolVersion. The current one is 1.0; a guide written for 0.3 or earlier describes a different version of the protocol.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}
5. Before connecting a third-party agent, review how it authorizes. The specification forbids treating the "authorization required" state as an authorization in itself: what gets authorized, and how, is defined by each implementation.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)}

Related concept: MCP

MCP connects an agent to its tools and data; A2A connects agents to each other. The A2A specification sums it up like this: MCP describes how an agent uses a specific capability, with structured inputs and outputs, and A2A describes how independent agents discover each other, split up tasks and exchange results as peers. An A2A client can hand a task to a remote agent that, to complete it, calls several tools over MCP.⁠^{[1]A2A Protocol, Linux Foundation. Agent2Agent (A2A) Protocol Specification (version 1.0.0)} The protocol's documentation puts it as an image: MCP is vertical, it goes deep into one agent; A2A is horizontal, it connects agents. Its example is a repair shop: the mechanic agent uses MCP to read the car's error codes and A2A to negotiate with a parts supplier's agent.⁠^{[7]A2A Protocol, Linux Foundation. A2A and MCP}

To go further, the agentic commerce reading path orders the rest of the protocols, and the entry on AP2 explains the extension of A2A that handles payment.