Hot AI News

Agent2Agent (A2A) protocol: How AI agents from different vendors hand work to each other

Agent2Agent (A2A) is an open standard that lets AI agents built by different companies find each other, share a task and return results, without knowing how the other agent works inside.

Agent2Agent (A2A) protocol: How AI agents from different vendors hand work to each other
Image: Amazon Web Services

The Agent2Agent protocol, or A2A, is an open standard that lets one AI agent ask another AI agent to do a piece of work. The two agents can come from different companies and run on different software. They only need to speak the same language about who they are, what they can do and how a job is going.

Google announced it on April 9, 2025. It now sits under the Linux Foundation.

A shared task board for software that never met

Picture a bank handling a disputed card payment. One agent checks the cardholder’s history. A second, built by a specialist vendor, tests the claim against chargeback rules. A third, written in-house, emails the customer for a missing receipt.

Without a standard, each pair of agents needs its own custom connection. With a standard, any agent can hand a job to any other, the way any browser can open any website.

That is the problem A2A addresses. It is not about making agents smarter. It is about making independent agents cooperate.

Advertisements

Cards, tasks and artifacts: The vocabulary of a handoff

A tablet and desktop monitor show abstract agent cards, task status panels and the same architectural photograph in two different interfaces.
Photo illustration: Aiholics

A2A defines a small set of building blocks. The names below come from the official specification at a2a-protocol.org.

  • Agent Card: a self-describing profile that an agent publishes. It lists the agent’s name, skills, supported interfaces and security requirements. Clients find it at a well-known address, /.well-known/agent-card.json, on the agent’s domain.
  • Message: one turn of conversation between a client agent and a remote agent. A message carries one or more parts, such as text or a file.
  • Task: the unit of work. The spec calls it the core unit of action. A task has an identifier and a status that changes over time.
  • Artifact: the output of a task, such as a finished report or a filled-in form.

The task status matters most. The spec lists states such as submitted, working, input required, auth required, completed, failed, canceled and rejected. A long job can therefore pause to ask a question, then resume, instead of failing or hanging.

Ordinary web plumbing underneath

A2A deliberately reuses familiar standards. The spec lists HTTP, JSON-RPC 2.0 and Server-Sent Events (a way for a server to stream updates over HTTP) as its guiding choices. The current specification also describes gRPC and plain REST-style bindings, and it defines the data model in Protocol Buffers.

Two features suit long jobs. Streaming lets a remote agent send progress updates as they happen. Push notifications let it call a webhook on the client’s side when a job that took hours is finally done. An agent advertises support for each in its Agent Card.

Security follows the same pragmatic line. The spec says A2A treats agents as standard enterprise applications, relying on established web security practices. Production traffic must be encrypted, and the server must authenticate every request. The card lists the accepted schemes, which include API keys, OAuth 2.0, OpenID Connect and mutual TLS. Cards can also carry digital signatures.

One detail is easy to miss. According to the spec, an agent must not treat the TASK_STATE_AUTH_REQUIRED state, by itself, as authorization. Asking for credentials mid-task does not grant them.

Advertisements

From a Google launch to a Linux Foundation project

Google Cloud launched A2A on April 9, 2025, with what it described as more than 50 technology partners. Named partners included Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow and Workday.

On June 23, 2025, Google Cloud donated the protocol to the Linux Foundation, which launched it as the Agent2Agent project. Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP and ServiceNow were the founding members. A Linux Foundation release in April 2026 reported version 1.0, the first stable specification, and more than 150 supporting organizations. The specification site shows 1.0.0 as the latest released version, with 0.3.0, 0.2.6 and 0.1.0 before it.

A2A is often mentioned beside the Model Context Protocol (MCP), which Anthropic introduced. Google’s launch post says A2A complements MCP. In short, MCP connects an agent to tools and data. A2A connects an agent to other agents that act as peers.

Where A2A already shows up in products

Amazon Connect is one place. Our report on Amazon Connect Resolutions noted that Amazon’s pricing page bills a step the same way whether a person, an AI agent or an API call performs it. It names a customer’s own agents, connected through the Resolutions A2A framework, among the agents that can do a step. A step in a refund or dispute case can therefore go to an agent the customer built elsewhere.

Amazon’s Connect documentation describes a related feature. A Connect AI agent can bring in another agent during a live contact. That agent can do work, talk to the customer, or both, then return control. External agents run on infrastructure their provider manages and connect over the A2A protocol through a WebSocket connection. Amazon says it added A2A extensions for voice streaming, session continuity and observability, because the base protocol does not provide them. The documentation describes this feature as API-only.

That last point shows how a standard works in practice. The core protocol covers discovery, tasks and results. Vendors fill gaps with extensions, which can erode some of the portability the standard promises.

Advertisements

What a standard does not solve

A2A tells agents how to exchange work. It does not decide whether the work is right. Some limits follow from that.

  • Trust: an Agent Card is a claim made by the agent’s owner. A client still has to decide which agents to believe and what data to share with them.
  • Accountability: when a task passes through several agents, someone must still be able to audit each action. Amazon’s own approach adds observability data on top of A2A for this reason.
  • Cost and complexity: every handoff is another step that may be billed, as the Resolutions pricing shows, and another place for an error to enter.
  • Extensions: vendor additions can make two A2A-compatible products work together less smoothly than the label suggests.

Safety research on multi-agent systems has begun to measure protocol-level risks in long chains of agents. Treat such work as early, and read any claim of seamless interoperability with that in mind.

What to watch next

Three things are worth following. First, whether the project’s roadmap items on registries, testing tools and security guidance mature, since those decide how easily a company can find and vet third-party agents. Second, how A2A and MCP are packaged together. The A2A site now carries a banner saying A2A joins the Agentic AI Foundation, though the details were not checked for this guide. Third, whether cloud vendors keep their extensions compatible, which will show if A2A becomes a real common layer or a label on separate ecosystems.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Daniel ReedAI Research, Safety & Ethics Analyst

Daniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like