04 Jul 2025 AI Engineering 5 min read

A2A Moves to the Linux Foundation

Google has donated the Agent2Agent (A2A) protocol to the Linux Foundation. AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow are founding participants. We examine what the protocol specifies, what it deliberately leaves out, and why neutral governance is the precondition for agent interoperability across organizational boundaries.

Agents Do Not Interoperate by Default

Every major software vendor now ships an agent platform. Copilot Studio, Agentforce, Google's Agent Development Kit, and dozens of open-source frameworks each define their own message formats, task lifecycles, and authentication schemes. An agent built on one platform cannot delegate work to an agent on another without custom integration code. With n platforms, that means up to n² point-to-point adapters. This is the enterprise integration problem of the 2000s, replayed with agents.

Protocols remove this quadratic cost. Anthropic's Model Context Protocol (MCP), released in November 2024, standardized how an agent reaches tools and data. The missing layer was agent-to-agent communication, especially across organizational boundaries. Google published Agent2Agent (A2A) in April 2025 to fill that gap. But a protocol controlled by a single vendor carries a governance risk that slows adoption. On June 23, 2025, that risk was removed.

An agent reaches its tools over MCP — one standard surface. 1/4

The Announcement in Facts

At Open Source Summit North America in Denver, the Linux Foundation announced the Agent2Agent project on June 23, 2025. Google transfers the A2A protocol specification, the accompanying SDKs, and developer tooling to the new entity. The founding participants are Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow. More than 100 companies support the protocol; AWS and Cisco are its newest validators. The code now lives under the a2aproject organization on GitHub.

Jim Zemlin, Executive Director of the Linux Foundation, framed the move as securing "long-term neutrality, collaboration and governance". The wording is standard for foundation announcements. The signal is not: five of the seven founding participants operate competing agent platforms. They are now committed to one wire format.

Project scope goes beyond the specification text. The transfer covers the SDKs, npm packages, sample implementations, and developer tooling that Google built since April. Under Linux Foundation rules, the project gets open technical governance: maintainers, public contribution paths, and vendor-neutral decision processes.

From Vendor Launch to Foundation Project

The timeline is unusually compressed for a standards effort. Google announced A2A on April 9, 2025 at Cloud Next, with more than 50 partners — Microsoft and AWS notably absent. Microsoft joined the working group on May 7 and committed A2A support for Azure AI Foundry and Copilot Studio. AWS and Cisco arrived with the foundation move. Eleven weeks passed between public launch and donation.

Speed is the point. Agent platforms are being deployed now, and every month without a common protocol produces more proprietary integrations that will have to be unwound later. Google stated in April that a production-ready version was planned for later in the year. Moving governance out first, before 1.0, reverses the usual order — and reduces the risk that the standard forks.

Date	Event
25 Nov 2024	Anthropic releases the Model Context Protocol (MCP)
9 Apr 2025	Google announces A2A at Cloud Next with 50+ partners
7 May 2025	Microsoft joins the A2A working group; support planned for Azure AI Foundry and Copilot Studio
23 Jun 2025	Linux Foundation launches the Agent2Agent project with seven founding participants

What the Protocol Specifies

The current specification is version 0.2. A2A uses JSON-RPC 2.0 over HTTPS as its transport. Every A2A server publishes an Agent Card: a JSON document describing identity, capabilities, skills, service endpoint, and authentication requirements. Clients use the card for discovery and to negotiate how to interact. Authentication reuses established schemes — OAuth 2.0, API keys, OpenID Connect — declared per agent rather than reinvented by the protocol.

Interaction is task-oriented. A client sends a message; the server creates a task with a defined lifecycle: submitted, working, input-required, completed, failed, or canceled. Outputs are typed artifacts. For incremental results, servers stream status and artifact updates over Server-Sent Events. For long-running work — hours or days — push notifications deliver updates to a client webhook. Messages carry typed parts, so agents can negotiate formats beyond plain text.

Why Neutral Governance Matters

A wire protocol is infrastructure. No company builds cross-organizational workflows on a format that one competitor can change unilaterally. Foundation ownership resolves exactly this: the specification, trademarks, and repositories belong to a neutral nonprofit, and changes go through open governance instead of a vendor roadmap. Intellectual property management and long-term stewardship are the explicitly stated goals of the transfer.

The pattern has precedent. Google donated Kubernetes to the newly formed CNCF in 2015; it became the default orchestration layer partly because competitors could adopt it without strategic exposure. OpenTelemetry followed the same path for observability. A2A now takes it for agent communication. AWS contributing to a Google-initiated protocol is the visible proof that neutrality changed the calculus.

What A2A Does Not Solve

A2A defines the conversation between agents, not the agents. It does not specify orchestration logic, reasoning quality, model selection, or how a task is actually executed. It does not replace MCP: MCP connects an agent vertically to tools and data, A2A connects agents horizontally as peers. Both layers are needed, and both sides state this explicitly.

Open problems remain. Cross-organizational identity federation is declared in the Agent Card but not solved by it; establishing trust between two companies' agents is still deployment work. There is no global discovery registry. Liability, billing, and audit semantics for delegated tasks are out of scope. And version 0.2 is not a finished standard — breaking changes before 1.0 are possible and likely.

Where This Goes Next

From where we stand in late June 2025, the near-term work is predictable: hardening the specification toward 1.0, additional transports, and identity. We expect the foundation move to unblock enterprise pilots first — legal and procurement teams object far less to a Linux Foundation standard than to a competitor's API. Cisco plans A2A support in its AGNTCY components; AWS has stated it intends to contribute. Concrete code from both would be the first test of the new governance.

The grounded vision: a procurement agent at one company negotiating delivery windows with a logistics agent at a supplier — different vendors, different clouds, one protocol. That requires MCP inside the organizational boundary and A2A across it. Whether A2A becomes the TCP/IP of agents or one protocol among several will be decided by implementations, not announcements. We are building against version 0.2 now and watching the project's governance as closely as its specification.