Contents
At a glance
- Category
- Agent concept - interoperability
- Full name
- Agent2Agent protocol
- Origin
- Introduced by Google in 2025, now under open governance at the Linux Foundation
- Discovery mechanism
- Agent cards describing identity, skills, endpoint and authentication
- Task model
- Long-running with streamed progress, clarification requests, and completion states
- Relationship to MCP
- Complementary - MCP reaches tools, A2A delegates to other agents
What A2A is
A2A, short for Agent2Agent, is an open protocol for communication between AI agents that were built by different people, on different stacks, possibly at different companies. It was introduced by Google in 2025 and moved to open governance under the Linux Foundation. Where MCP standardises how an agent reaches tools and data, A2A standardises how one agent asks another agent to do something.
The distinction is the point. A tool is a function with known inputs and outputs. Another agent is an opaque, autonomous system with its own instructions, memory and context that it will not share. Delegating to it is a negotiation - what can you do, here is the task, tell me when it is done - rather than a function call.
How agents find and use each other
Each agent publishes an agent card: a machine-readable description of its identity, its skills, its endpoint and its authentication requirements. A client agent reads the card, decides the remote agent is suitable, and sends a task. Tasks have a lifecycle rather than a single response, because real delegated work takes time: they can be accepted, worked on with progress updates streamed back, require additional input, and eventually complete or fail. Exchanges carry structured parts, so agents can pass files and data rather than only prose.
That long-running, streamed task model is the substantive difference from a request-response API. An agent asked to prepare a quote might come back in four minutes having asked one clarifying question along the way, and the protocol has a shape for that.
Where it stands today
A2A is considerably younger and less widely deployed than MCP, and the realistic assessment as of August 2026 is that it is a standard to watch rather than a dependency to build a small business on. The interesting scenario it enables - your booking agent negotiating directly with a supplier's scheduling agent - requires the other side to have adopted it too, and cross-organisation agent delegation raises questions about liability, audit and authorisation that the protocol defines mechanisms for but that contracts have to settle.
Inside one organisation, A2A can be used to connect agents owned by different teams on different stacks without forcing a common framework. That is a real use, though many teams find a multi-agent framework simpler when everything is under one roof.
What you should know
A2A and MCP are complementary, not competing: an agent typically uses MCP to reach its own tools and A2A to delegate to peers. If someone presents them as alternatives, that is a sign to read the specifications rather than the summary.
For most businesses today the sensible position is awareness. Build your agents with clean boundaries and explicit interfaces - the design discipline A2A encodes - so that adopting it later is an adapter rather than a rewrite. And apply the same guardrails you would to any external system: an agent you do not control asking your agent to take an action is an untrusted request, with all the injection implications that carries.
A2A FAQ
What is the A2A protocol?
An open standard that lets AI agents built by different parties advertise what they can do and delegate tasks to each other. Each agent publishes an agent card describing its skills and endpoint, and clients send long-running tasks rather than single request-response calls.
What is the difference between A2A and MCP?
MCP connects an agent to tools and data - functions with defined inputs and outputs. A2A connects an agent to another agent, which is an autonomous system with its own context that it will not expose. They are complementary and most serious agent stacks will eventually use both.
What is an agent card?
A machine-readable description an agent publishes about itself: who it is, what skills it offers, where to reach it, and how to authenticate. It is the discovery mechanism, so a client agent can decide whether a remote agent is suitable before sending it any work.
Should my business adopt A2A now?
Probably not as a dependency. As of August 2026 it is younger and far less deployed than MCP, and its main promise requires the other party to have adopted it too. Build agents with clean interfaces so adopting it later is an adapter rather than a rewrite.
Can A2A be used inside one company?
Yes - it lets agents owned by different teams on different stacks talk without standardising on one framework. Where everything is under one roof and one team, a multi-agent framework is usually simpler and gives you shared state, which A2A deliberately does not.
Is it safe to let another company's agent call mine?
Only with the same care you would apply to any external caller: authenticate it, authorise specific actions, rate limit it, log everything, and treat its requests as untrusted input subject to prompt injection. The protocol provides mechanisms; liability and audit are contract questions.
Primary source
Checked August 2026. Anything priced or rate-limited moves, so verify against the source before you budget.
Want this built rather than explained?
Knowing what a2a means is one thing; having it running in your business is another. We scope one workflow at a time, quote a fixed price in writing after a free consult, build it, and hand it over with documentation you own.
