Glossary
What is A2A Protocol (Agent2Agent)?
The Agent2Agent (A2A) Protocol is an open standard, announced by Google in April 2025 and later contributed to the Linux Foundation, that lets AI agents built on different frameworks and by different vendors discover each other, exchange messages securely, and collaborate on tasks — regardless of the underlying model or stack.
Last updated:
A2A addresses a practical problem: as companies deploy AI agents, those agents need a common language to work with agents outside their own vendor ecosystem. The protocol defines how an agent publishes an 'Agent Card' describing its capabilities, how another agent initiates a task, and how the two exchange messages and artifacts over standard web transports (HTTP, JSON-RPC, Server-Sent Events).
A2A is complementary to the Model Context Protocol (MCP): MCP connects a model to tools and data sources, while A2A connects agents to other agents. A useful mental model — MCP is how an agent uses its hands; A2A is how it talks to strangers.
For B2B platforms like AgentDoor, open agent-interoperability standards matter because buyer and vendor agents are built by different organizations and must still hold a structured, auditable conversation. AgentDoor's 'door' concept is a business-layer application of the same idea: a published, permissioned endpoint where outside agents can ask questions and check fit.
How it works
Three pieces carry most of the protocol.
The **Agent Card** is a document an agent publishes about itself: what it can do, where to reach it, and how to authenticate. It is the discovery step — the reason a client agent can approach a server agent it has never met without a pre-arranged integration.
A **task** is the unit of work. A client agent opens one, the server agent works on it, and the task carries a lifecycle rather than a single request/response: it can stay open across several turns, ask the client for more input, and report progress while it runs. Long-running work is the normal case, not an edge case, which is why streaming over Server-Sent Events is part of the design rather than bolted on.
**Messages and artifacts** are what move through the task. Messages are the conversational turns; artifacts are the outputs the task produces. Both are structured, and both can carry more than text — which matters for procurement, where the answer to a question is often a document rather than a sentence.
Authentication is deliberately not reinvented: the Agent Card declares which standard web auth schemes the agent accepts, and the caller uses one of them.
A2A and MCP, side by side
The two are asked about together constantly, so it is worth being precise. MCP standardises the link between a model and the tools and data it uses — an inward-facing connection, usually inside one organisation's trust boundary. A2A standardises the link between two agents that may belong to different companies, run on different stacks, and have no shared infrastructure at all.
That difference in trust boundary is the real distinction. An MCP tool call happens inside a system you control. An A2A exchange crosses an organisational line, which is why discovery, declared capabilities, authentication and an auditable task record are all first-class in A2A and largely absent from MCP. A single agent typically speaks both.
Where the standard actually is today
The specification is public, and governance sits with the Linux Foundation rather than with a single vendor — which is the part that matters for anyone deciding whether to build against it. Ecosystem support has grown quickly since the April 2025 announcement.
What has not arrived at the same pace is traffic. Cross-organisation agent conversations are still rare in production, and the reason is rarely the protocol. It is that most companies have nothing on the other side of the connection: no approved, current material an outside agent could be safely allowed to query, and no governance over what an agent may say on the company's behalf.
What a company can do now
The useful preparation is not protocol work. It is deciding what an outside agent would be allowed to ask, which source material is approved and current enough to answer from, and who reviews what goes out. A vendor that has settled those three questions can be reached over A2A, or over whatever standard wins, with modest engineering. A vendor that has not will still be answering the same security questionnaire by hand in three years, whichever protocols ship.
Frequently asked questions
- What does A2A stand for?
- In the context of AI agents, A2A stands for Agent2Agent — the open protocol announced by Google in April 2025 and now governed by the Linux Foundation. Note that the same abbreviation is used in payments for account-to-account transfers, which is unrelated to this protocol.
- What is the difference between A2A and MCP?
- MCP (Model Context Protocol) standardizes how an AI model connects to tools and data. A2A standardizes how independent agents communicate with each other. An agent may use MCP internally for its tools and A2A externally to talk to other agents.
- Does a vendor need to implement the A2A protocol to be reachable by buyer agents?
- Not necessarily. The protocol standardises the transport layer, but what most vendors lack is the business layer above it: a governed endpoint that answers outside agents' questions from approved material only. Platforms like AgentDoor provide that layer without the vendor writing any agent code.
- Is the A2A protocol production-ready?
- It is early. The specification is public and governed by the Linux Foundation, and ecosystem support is growing, but cross-organisation agent traffic is still rare in production. That is precisely why the space is open: the standards exist before most companies have anything answerable behind them.
- Who created the A2A protocol, and who controls it now?
- Google announced it in April 2025 and subsequently contributed it to the Linux Foundation, which now hosts the project. That hand-off is the reason it can be treated as a neutral standard rather than one vendor's interface.
- What is an Agent Card?
- A document an agent publishes describing what it can do, where to reach it, and which authentication schemes it accepts. It is what makes discovery possible: another agent can read the card and start a task without any pre-arranged integration between the two organisations.
- Is A2A the same as agentic commerce or agent payments protocols?
- No. A2A defines how agents communicate; it does not define how a purchase is authorised or settled. Payment-oriented protocols address that separate layer, and an end-to-end agent transaction would typically involve both.
Skip the discovery calls.
Describe what you need and let AgentDoor's agents interview the vendors for you — you get a decision-ready shortlist, not six meetings.
We'll reach out at launch. No spam, ever.