Agent-to-agent protocols (A2A)
How agents built by different teams on different frameworks discover each other and hand work back and forth, and how that differs from MCP.
What problem it solves
MCP connects an agent to tools and data. A2A, the Agent2Agent protocol, addresses a different gap: letting agents built by different companies, on different frameworks, discover one another and hand work back and forth1. Before it, connecting an agent in one company’s stack to an agent in another’s needed custom integration code for every pairing1. Google announced it in April 2025 and handed it to the Linux Foundation that June1.
Diagram: my own.
The core idea: opaque agents
The key design choice is that the agent on the other end is opaque. You do not get its tools, memory, model or internal plan; you get a card, a task ID and the states the task passes through2. The repository describes it as communication between opaque agentic applications, as agents and not as tools3. That is what lets organisations cooperate without exposing how their agents work.
Four ideas
One builder’s guide says A2A standardises four things, with the rest being transport detail2:
- Agent Card: a JSON manifest saying who the agent is, what it can do, where to reach it and how to authenticate2. It is typically published at a well-known URL on the agent’s domain, so discovery needs only the domain1.
- Task: the unit of work with an ID and a lifecycle.
- Message: the turns exchanged while working on a task.
- Artifact: the outputs the task produces.
Diagram: my own, after the flow described in 1 and 2.
Task lifecycle
A request does not have to be a single synchronous call. The remote agent can answer immediately or return a task the client can track until it finishes1. A task moves through states, roughly:
Diagram: my own, a simplified view; check the specification for the exact state names and transitions.
The “input required” state is how a remote agent asks the client for more information mid-task, so long-running, multi-turn work fits naturally.
Governance and status
A2A moved to the Linux Foundation in June 2025, with support from more than 100 technology vendors named in coverage, including AWS, Microsoft, Cisco, Salesforce and SAP4. Versions and extension mechanisms have continued to evolve, and a v1.0 builder’s guide was published in August 20262. Because the spec is moving, treat details here as a sketch and read the current specification before implementing.
When you need it
Most agent systems today are one agent plus tools. You reach for A2A when the other party is genuinely a separate agent you do not control: a vendor’s agent, another team’s service, a partner organisation. If you own both sides, subagents inside one harness are simpler and give you more control over context. A2A trades that control for interoperability. Whatever you choose, remember that a remote agent’s replies are untrusted input; see Safety and guardrails.
Sources
- [1]Agent2Agent Protocol (A2A) explained — PPC Land · 2025PARAPHRASED
- [2]A2A v1.0: a builder's guide, part 1 — discovery, tasks, and clients — Rohit Ghumare · Agentic AI Foundation · 2026PARAPHRASED
- [3]A2A protocol (repository and specification) — a2aproject on GitHubPARAPHRASED
- [4]Google donates A2A protocol to Linux Foundation — Techstrong.ai · 2025PARAPHRASED