Protocols & news

What is A2A (Agent2Agent)? How AI agents talk to each other

Listen · 10 min

Narrated by an AI voice. Code examples are on the page.

Short answer

A2A (Agent2Agent) is an open protocol that lets AI agents from different companies find each other, hand off work and send back results. One agent reads another agent's agent card, sends it a message, and gets back a task that moves through set states until it is done. Google created A2A in 2025, it reached a stable v1.0 in 2026, and it now sits in the Agentic AI Foundation under the Linux Foundation.

Key takeaways

  • A2A connects agents to agents; MCP connects agents to tools.
  • Work is tracked as tasks that can pause for input or approval.
  • Version 1.0 added signed agent cards and three equal bindings.

What is A2A (Agent2Agent)?

A2A, short for Agent2Agent, is an open protocol that lets one AI agent find another agent, give it a job and get the result back, even when the two were built by different companies on different software. The A2A project's GitHub page describes it as "an open protocol enabling communication and interoperability between opaque agentic applications."

"Opaque" is the important word. A2A does not ask agents to share how they work inside. The specification says agents collaborate "without needing to share their internal thoughts, plans, or tool implementations." Each agent declares what it can do, and the other agent works with that.

If you only want the one-line version, our A2A glossary entry has it, and the agent card entry covers the discovery file. This article goes one level deeper: how a conversation between two agents actually runs, how A2A handles approvals, and how it has changed.

How does A2A work, step by step?

An A2A exchange has three moves: read the other agent's card, send it a message, and follow the task it creates until it finishes.

The specification names the two sides. An A2A client is "an application or agent that initiates requests." An A2A server, also called a remote agent, is "an agent or agentic system that exposes an A2A-compliant endpoint."

  1. Discovery. The client reads the server's agent card, "a JSON metadata document published by an A2A Server, describing its identity, capabilities, skills, service endpoint, and authentication requirements." A common place for it is /.well-known/agent-card.json.
  2. Message. The client sends a message. A message is "a communication turn between a client and a remote agent," made of one or more parts. A part can hold text, a file reference or structured data.
  3. Task. The server either answers directly with a message, for simple requests, or creates a task. The specification calls a task "the fundamental unit of work managed by A2A." Results come back as artifacts, such as a document or structured data.

Here is a shortened version of the specification's example over the HTTP binding, with a weather question:

POST /message:send HTTP/1.1
Host: agent.example.com
Content-Type: application/a2a+json

{
  "message": {
    "role": "ROLE_USER",
    "parts": [{"text": "What is the weather today?"}],
    "messageId": "msg-uuid"
  }
}

The reply is a task with an ID, a status of TASK_STATE_COMPLETED and an artifact holding the answer. A related context ID can group several tasks into one longer conversation.

The same operations run over three bindings: JSON-RPC, gRPC and HTTP+JSON (REST). The core operations include SendMessage, SendStreamingMessage, GetTask, ListTasks and CancelTask.

What happens when an agent needs more input or a person's approval?

A2A lets a task pause and wait, rather than fail. This matters for anything involving money or a shopper's account.

Every task moves through defined states. The main ones are submitted, working, completed, failed, canceled and rejected. Two states are pauses:

  • Input required. The specification says agents "can request additional input mid-processing by transitioning a task to the input-required state." The client answers with a new message on the same task.
  • Auth required. The specification's examples include "an agent requiring human approval before a destructive action is taken." The agent moves the task to TASK_STATE_AUTH_REQUIRED and the client deals with it, for example by contacting "another human, agent, or service."

A2A also allows for work that takes a long time. The specification lists support for "long-running tasks and interactions that may involve human-in-the-loop scenarios." A client can stream updates, poll with GetTask, or register a webhook to receive push notifications.

For a store, this is the shape you would expect: a shopper's agent asks for something, your side can ask a follow-up question or hold the request for a person, and the task resumes when that is settled.

How has A2A changed since 2025?

A2A went from a Google project to a stable, foundation-governed standard in less than a year and a half. These are the steps we could confirm from primary sources:

  • April 2025. Google launched A2A. The Linux Foundation later described it as "a collaborative effort launched by Google in April."
  • 23 June 2025. The Linux Foundation announced the A2A project, describing it as "an open protocol created by Google for secure agent-to-agent communication and collaboration."
  • 12 March 2026. Version 1.0.0 was released on GitHub.
  • 9 April 2026. The Linux Foundation said more than 150 organizations supported A2A and called v1.0 "its first stable specification."
  • 27 August 2026. The A2A project said it had "officially been accepted as a Growth Stage project at the Agentic AI Foundation (AAIF)," alongside MCP, goose and AGENTS.md. See what the AAIF is.

Version 1.0 changed a lot under the hood. The project's change notes list "Agent Card signature verification using JWS and JSON Canonicalization," "Formal specification of all three protocol bindings with equivalence guarantees," multi-tenancy support, and a new ListTasks operation. Some changes broke older clients, such as task states moving to upper-case names like TASK_STATE_COMPLETED.

Signed agent cards are the change that matters most for trust. The Linux Foundation says they provide "cryptographic identity verification," so a client can check that a card was not altered.

Note. A2A is still evolving. The Linux Foundation lists an interoperability specification and registry work on the roadmap. Check a2a-protocol.org for the current version before you build.

How does A2A relate to MCP, AP2 and UCP?

A2A is the agent-to-agent layer. Other protocols plug into it or sit beside it.

MCP. The specification calls A2A and MCP "complementary protocols." It describes MCP as the "how-to" for an agent to use a tool, and A2A as "how agents partner or delegate work." The A2A project's own image: MCP is the "vertical integration layer" and A2A is "the horizontal protocol." A server agent may well use MCP behind the scenes to finish an A2A task. See what MCP is.

AP2. Google said when it announced AP2 in September 2025 that it "can be used as an extension of the Agent2Agent (A2A) protocol and Model Context Protocol (MCP)." The same post introduced an A2A x402 extension for crypto payments.

UCP. Google's checkout standard lists A2A as one of its transports, next to REST, MCP and an embedded option. In a UCP business profile, the specification says the "endpoint for A2A transport refers to the Agent Card URL for the agent." See what UCP is.

For a side-by-side table of all three, read MCP vs A2A vs UCP.

Does your store need A2A?

Not on its own, at least not yet. Most store owners will meet A2A indirectly: through a platform, through UCP's A2A transport, or by publishing an agent card that other agents can read.

The adoption so far is mostly in enterprise software. The Linux Foundation names supply chain, financial services, insurance and IT operations. The A2A project names ServiceNow, Salesforce, Atlassian and SAP among the platforms using it.

What is worth doing now:

  1. Publish an accurate agent card, so any agent that looks can see what your business offers.
  2. Keep it honest. A card that lists skills you cannot actually perform will send agents to dead ends.
  3. Make sure anything that spends money pauses for your rules or a person, which A2A's task states allow for.

How Sill helps. Sill publishes a signed, A2A-compatible agent card for each verified site and keeps it current, and it supports A2A message/send for your skills. When an AI agent places an order, your rules approve it, decline it or hold it for a person to review, and every decision is signed and recorded. Run a free readiness scan to see what agents find on your site today.

Questions people ask

Is A2A free to use?

Yes. The specification and core repository are published under the Apache 2.0 licence, so anyone can implement it without paying a fee.

Which platforms support A2A?

The Linux Foundation says Microsoft integrated A2A into Azure AI Foundry and Copilot Studio, and AWS added support through Amazon Bedrock AgentCore Runtime. The A2A project also names Google Cloud, and frameworks such as LangGraph and CrewAI.

Can an A2A agent see how another agent works inside?

No, and that is by design. The specification says agents collaborate "without needing to share their internal thoughts, plans, or tool implementations." Each agent only sees what the other one declares and sends back.

What programming languages have A2A SDKs?

The Linux Foundation says the SDKs grew from a single Python implementation to five production-ready languages, including JavaScript, Java, Go and .NET.

Sources

  1. A2A Protocol: Specification (v1.0) · checked 17 Nov 2026
  2. A2A Protocol: What's new in v1.0 · checked 17 Nov 2026
  3. A2A Protocol blog: A new chapter for A2A, joining the Agentic AI Foundation · checked 17 Nov 2026
  4. Linux Foundation: Linux Foundation launches the Agent2Agent protocol project · checked 17 Nov 2026
  5. Linux Foundation: A2A Protocol surpasses 150 organizations in first year · checked 17 Nov 2026
  6. GitHub: a2aproject/A2A · checked 17 Nov 2026
  7. GitHub: a2aproject/A2A releases · checked 17 Nov 2026
  8. Agentic AI Foundation · checked 17 Nov 2026
  9. Google Cloud: Announcing Agent Payments Protocol (AP2) · checked 17 Nov 2026
  10. AP2: Agent Payments Protocol · checked 17 Nov 2026
  11. UCP specification overview · checked 17 Nov 2026

Written by The Sill Signals Team. Last reviewed 17 Nov 2026.

Share this articleLinkedInXBlueskyEmail

Keep reading

See which AI agents visit your site, and decide what they may do.

Scan your site free