The Agent2Agent Protocol (A2A) is an open standard that lets artificial intelligence (AI) agents built by different companies, on different software frameworks, discover one another and hand work back and forth. One agent publishes a machine-readable description of what it can do, called an Agent Card. Another agent reads that card, sends a message and receives either an immediate answer or a task that it can track until completion. Google announced A2A on April 9, 2025, and handed it to the Linux Foundation that June. Before A2A, connecting an agent built in one company's stack to an agent in another's required custom integration code for every pairing. As of November 2026 the stable version is 1.0.1, and version 1.1 is at release-candidate stage.

How A2A works

A2A describes a conversation between two roles. The client is the agent, or the application acting for a user, that wants something done. The remote agent, also called the A2A server, is the agent that does it. The remote agent is opaque: the client sees only what it chooses to return, not its reasoning, memory or tools.

An exchange runs in roughly five steps.

Discovery. The client fetches the remote agent's Agent Card. The default location is a well-known web address, https://{domain}/.well-known/agent-card.json, following RFC 8615, the internet standard for such paths. The card can also be distributed through a private registry or configured directly. The specification does not define a standard registry interface.

Reading the card. The Agent Card is a JSON document. Its main fields are name, description and provider; supportedInterfaces, which lists each endpoint URL with its transport and protocol version in order of preference; capabilities, which flags support for streaming and push notifications; securitySchemes and securityRequirements, which state how a client must authenticate; and skills, an array describing each task the agent offers, with an id, a description, tags, example prompts and accepted input and output formats.

Sending a message. The client authenticates using one of the declared schemes (API key, HTTP authentication, OAuth 2.0, OpenID Connect or mutual TLS) and calls SendMessage. A message carries a role (user or agent) and one or more parts, each holding text, a file or structured data.

Task lifecycle. The remote agent can reply with a direct message or create a task, a stateful unit of work with its own identifier. Tasks move through defined states: SUBMITTED, WORKING, INPUT_REQUIRED, AUTH_REQUIRED, COMPLETED, FAILED, CANCELED and REJECTED. The last four are terminal and accept no further messages. The two "required" states pause a task until the client supplies more information or credentials, which is how A2A handles a mid-job negotiation or approval. A contextId groups several related tasks and messages into one conversation.

Results. Outputs arrive as artifacts, such as a document, an image or a JSON payload. Clients can wait for a blocking response, poll with GetTask, open a stream through SendStreamingMessage or SubscribeToTask (delivered as server-sent events), or register a webhook for push notifications.

Version 1.0 defines eleven operations, among them ListTasks, CancelTask and GetExtendedAgentCard, which returns a fuller card to authenticated clients. Each operation is mapped onto three protocol bindings: JSON-RPC 2.0 over HTTP, gRPC, and plain HTTP with JSON. Clients state the version they speak in an A2A-Version header; a server that does not support it returns a VersionNotSupportedError.

Who operates it

A2A has no central operator. The project's August 2026 announcement named Google Cloud, Amazon Web Services' Bedrock AgentCore Runtime and Microsoft's Azure AI Foundry as offering native support, alongside ServiceNow, Salesforce, Atlassian and SAP. Framework developers, including LangGraph, CrewAI, Pydantic AI, AG2 and IBM's BeeAI, ship A2A clients and servers.

In advertising, the operators are the companies building buying and selling agents. On the sell side, a publisher or supply-side platform (SSP) would expose a seller agent whose card lists skills such as answering a deal proposal. On the buy side, an agency or demand-side platform (DSP) runs a buyer agent that reads those cards and opens tasks. IAB Tech Lab's reference agents use the Model Context Protocol (MCP) as the primary route and A2A messages over JSON-RPC 2.0 for conversational queries.

Origin and evolution

The first commit to Google's A2A repository is dated March 25, 2025. Google announced the protocol publicly on April 9, 2025, during its Cloud Next conference, in a developer blog post by Rao Surapaneni, Miku Jha, Michael Vakoc and Todd Segal that cited "more than 50 technology partners", including Atlassian, PayPal, Salesforce and SAP. The post set out five design principles: embrace agentic capabilities, build on existing standards, be secure by default, support long-running tasks, and be modality agnostic.

The first tagged release, v0.1.0, followed on May 19, 2025. On June 23, 2025, at Open Source Summit North America in Denver, the Linux Foundation announced it would host the project, citing support from "more than 100 leading technology companies". "We are happy to be the new home of the Agent2Agent Protocol project," Jim Zemlin, the foundation's executive director, said.

Version 0.3.0, released at the end of July 2025, made several breaking changes. It moved the card from agent.json to agent-card.json, added cryptographic signatures to the Agent Card, added mutual TLS as a security scheme and introduced the extended card. IBM's Agent Communication Protocol was folded into A2A in August 2025, according to AI Magazine.

A release candidate for version 1.0 appeared on January 29, 2026, and v1.0.0 shipped on March 12, 2026, described by the project as its first stable, production-ready version. It separated abstract operations from transport mappings, added ListTasks and multi-tenancy, dropped the OAuth 2.0 implicit and password flows, and defined signed Agent Cards using JSON Web Signature (RFC 7515). Interaction behaviour broke compatibility with 0.3, but the Agent Card can advertise both versions at once so clients migrate gradually. Version 1.0.1 followed on May 26, 2026.

Governance moved again in August 2026, when A2A joined the Agentic AI Foundation (AAIF), the Linux Foundation body formed in December 2025 around Anthropic's donated MCP, Block's goose and OpenAI's AGENTS.md. Sources disagree on the details. The AAIF's own post is dated August 17, 2026 and describes A2A as a "hosted project"; the A2A project's blog post is dated August 27 and calls it a "Growth Stage project". Both cite backing from more than 150 organisations. "As organizations move multi-platform agentic systems into production, interoperability is becoming essential," Surapaneni said in the AAIF announcement.

The project's Technical Steering Committee had eight member companies at the 1.0 launch. Bloomberg became the ninth on October 20, 2026, joining Google, Microsoft, Cisco, AWS, Salesforce, ServiceNow, SAP and IBM. A v1.1.0 release candidate was tagged on October 29, 2026. It adds SendLiveMessage for two-way streaming while a task runs, lets clients write artifacts and defines a media type, application/a2a-agent-card+json. On October 22, documentation began steering discovery towards AI Catalog registries, since the well-known address holds only one card per domain.

Why it matters for advertising

Advertising standards have adopted A2A as a transport rather than competing with it. When the Ad Context Protocol launched in October 2025, it offered both MCP and A2A access, with A2A positioned for multi-agent workflows and approval chains. On January 6, 2026, IAB Tech Lab's agentic roadmap proposed extending OpenRTB, AdCOM, OpenDirect and VAST with MCP, A2A and gRPC instead of writing new protocols; that work was later named AAMP. When IAB Tech Lab's chief operating officer listed the 13 functions on which AAMP and AdCP overlap, "MCP and A2A communication" was one of them.

Commerce is the other channel. Google's Universal Commerce Protocol, launched in January 2026, runs over MCP, A2A or REST, so a merchant's agent can negotiate with a shopping assistant.

Limitations and disputes

Discovery and trust remain thin. The specification offers no standard registry API, and it states that a signature on an Agent Card does not by itself establish trust. Whose keys to believe is left to deployments.

Privacy signals are absent. An October 2026 academic audit covered by PPC Land found that neither MCP nor A2A carries a privacy field, so a user's opt-out of advertising data sharing cannot travel with a delegated task.

Adoption is uneven. The same audit found that of 22 AI companies, 16 used MCP and five used A2A in their own products, and that Anthropic, OpenAI, xAI and DeepSeek did not use A2A in theirs. Supporter counts are self-reported and do not measure production traffic. In advertising, Adform's chief technology officer said in September 2026 that little money flows through either of ad tech's agent protocols, the layer that would sit on top of A2A.

Version churn has been heavy: breaking changes in 0.3 and again in 1.0 within a year of launch.

What A2A is not

MCP. The Model Context Protocol connects one agent to tools and data sources, such as a database or an API. The A2A project calls MCP "vertical" and A2A "horizontal"; many systems use MCP inside an agent and A2A between agents.

AdCP and AAMP. These are advertising vocabularies, defining operations such as fetching products or creating a media buy. They specify what advertising agents say; A2A is one of the channels that carries it.

AP2 and UCP. Google's Agent Payments Protocol and Universal Commerce Protocol add payment and checkout semantics on top of A2A or MCP.

A2A payments. In banking, "A2A" usually means account-to-account transfers, an unrelated use of the same abbreviation.

Recent developments

The past three months brought the AAIF move, Bloomberg's committee seat, catalog-based discovery and the v1.1.0 release candidate. PPC Land's October 31 report on the AI privacy audit argued that emerging advertising protocols will inherit whatever MCP and A2A carry, or fail to carry, making the absence of a privacy signal in A2A a question for ad tech as much as for agent developers.

Timeline

  • March 25, 2025 - First commit in Google's A2A repository.
  • April 9, 2025 - Google announces A2A at Cloud Next with more than 50 partners.
  • May 19, 2025 - First tagged release, v0.1.0; v0.2.0 follows on May 21.
  • June 23, 2025 - Linux Foundation announces it will host the A2A project.
  • July 30-31, 2025 - v0.3.0 renames the card path to agent-card.json and adds card signatures and mutual TLS.
  • August 2025 - IBM's Agent Communication Protocol merges into A2A.
  • October 15, 2025 - Ad Context Protocol launches with MCP and A2A access.
  • December 9, 2025 - Agentic AI Foundation formed around MCP, goose and AGENTS.md.
  • January 6, 2026 - IAB Tech Lab roadmap extends ad standards with MCP, A2A and gRPC.
  • January 11, 2026 - Google launches the Universal Commerce Protocol, which runs over A2A among other transports.
  • January 29, 2026 - v1.0.0 release candidate.
  • March 12, 2026 - v1.0.0, the first stable release, with signed Agent Cards and multi-tenancy.
  • May 26, 2026 - v1.0.1.
  • August 17-27, 2026 - A2A joins the Agentic AI Foundation (sources differ on the date).
  • October 20, 2026 - Bloomberg joins the Technical Steering Committee as its ninth member.
  • October 22, 2026 - Documentation shifts discovery towards AI Catalog registries.
  • October 29, 2026 - v1.1.0 release candidate adds SendLiveMessage.

Summary

Who: Google created A2A and announced it in April 2025. The Linux Foundation has hosted it since June 2025, and since August 2026 it sits within the Agentic AI Foundation, governed by a nine-company steering committee of Google, Microsoft, Cisco, AWS, Salesforce, ServiceNow, SAP, IBM and Bloomberg.

What: An open protocol through which AI agents publish Agent Cards describing their skills, then exchange messages, tasks and artifacts over JSON-RPC 2.0, gRPC or HTTP with JSON.

When: Announced April 9, 2025; v0.3.0 in July 2025; stable v1.0.0 on March 12, 2026; v1.0.1 on May 26, 2026; v1.1.0 release candidate on October 29, 2026.

Where: In enterprise agent platforms and clouds, including Google Cloud, AWS and Azure, and in advertising as a transport for AdCP, IAB Tech Lab's AAMP agents and Google's commerce protocols.

Why: Agents built on different frameworks by different companies could not delegate work to each other without custom integration. A2A provides a shared way to find an agent, authenticate, hand it a task and track the result, though trust, privacy signals and real-world adoption remain open questions.