Clarion Street

A2A Joins the Agentic AI Foundation — What Protocol Consolidation Actually Changes (and What It Doesn't)

·6 min read

On August 17, 2026, the Agent2Agent protocol became a hosted project of the Agentic AI Foundation (AAIF), the Linux Foundation body that already stewards Anthropic’s Model Context Protocol, Block’s goose runtime, OpenAI’s AGENTS.md convention, and the agentgateway proxy. The move puts the two most prominent open interoperability protocols in the agent stack under one neutral roof: MCP, which standardizes how agents reach tools and data, and A2A, which standardizes how agents discover and delegate to each other.

Most coverage treats this as a handover of A2A from Google to the foundation. That framing is wrong in a way that matters. Google launched A2A in April 2025 and donated the specification, SDKs, and tooling to a Linux Foundation project in June 2025, alongside AWS, Cisco, Microsoft, Salesforce, SAP, and ServiceNow. A2A has been under Linux Foundation neutral governance for about a year. What changed on August 17 is which foundation hosts the work. AAIF itself only formed on December 9, 2025, anchored by MCP, goose, and AGENTS.md. This is a consolidation of two already-neutral projects under one newer, more focused umbrella. A hosting clarification, not a control transfer.

Two protocols, different jobs

MCP and A2A do different jobs, and they complement each other rather than compete. MCP is the agent-to-tool layer. A2A is the agent-to-agent layer. Both need to exist for a multi-agent system to work, and keeping them under the same foundation means the two specifications can now coordinate on the places where they touch: how an agent that speaks MCP discovers another agent that speaks A2A, and how identity and authorization are handled across both.

What neutral governance actually buys

The real benefit of the move is the removal of single-vendor capture. A2A no longer turns on Google’s roadmap, and MCP no longer turns on Anthropic’s. A neutral foundation that hosts both gives them a single venue to coordinate, which neither had when they lived under separate Linux Foundation projects. That is a modest but genuine improvement, and it is the main thing the announcement actually changes.

What consolidation does not standardize

Putting MCP and A2A in the same foundation does not standardize enterprise authorization policy. The A2A spec has normative requirements: servers must authenticate every request and enforce per-skill and per-data authorization under least privilege. But the spec is explicit that the authorization logic itself “is specific to the agent’s implementation, the data it handles, and applicable enterprise policies.” The protocol tells you to enforce authorization. It does not tell you what to enforce.

Identity is handled the same way. Authentication is delegated to HTTP-layer standards, OAuth2 and OIDC or API keys, and credentials are acquired out of band. The payload itself carries no identity or credential material. That means the protocol does not solve key-to-organization binding or delegated authority. A signed agent card is only as trustworthy as the recipient’s ability to confirm that the signing key actually belongs to the organization named on the card, and the spec leaves that question open.

Google’s own June 2025 “what’s next” list is the honest tell. Over a year ago it named trustworthy agent identity, delegated agent authority, governance policy, agent security, and reputation as open topics. They are still open. The hard problems remain the deployer’s burden.

The stack is still fragmented

Consolidating A2A and MCP does not mean the agent interoperability stack has converged. The adjacent commerce protocols are still multiplying. Google Cloud and PayPal are extending A2A into agentic commerce through the Agent Payments Protocol (AP2). OpenAI has contributed a separate Agent Commerce Protocol (ACP) to the AAIF ecosystem. Cisco’s AGNTCY project still overlaps A2A on discovery and identity. So the moment A2A and MCP come under one roof, the surrounding stack is still MCP, A2A, AP2, ACP, and AGNTCY, with overlapping scope in at least two places.

Whether that gets better or worse under consensus governance is an open question. Neutral governance removes single-vendor capture, which is the real benefit, but it also adds coordination overhead. AAIF is a young foundation, formed in December 2025, and it now has to coordinate A2A’s release cadence, MCP’s, and a growing set of adjacent protocols. I would not bet that consolidation speeds anything up. It might be the venue where the commerce protocols eventually merge, or it might just be a cleaner place to watch them fork.

The 150+ figure needs a caveat

A2A is backed by more than 150 supporting organizations and is deployed in production at OS scale, including Huawei’s Celia assistant on HarmonyOS and Tencent’s WeChat, plus Google Cloud, Azure AI Foundry, and AWS Bedrock AgentCore. A2A v1.0 shipped in March 2026 with multi-protocol bindings, version negotiation, multi-tenancy, and cryptographically signed agent cards.

The “150+” number comes from the foundation and its vendors, not from independent measurement. Hosting an A2A endpoint is not the same as supporting the full surface: bindings, version negotiation, multi-tenancy, and signed cards. Conformance depth varies widely, and there is no independent registry of certified A2A implementations to check against.

That leaves buyers with a checklist the announcement does not answer. Which A2A revision and bindings does a given vendor actually implement? Does it negotiate versions with older clients? Does it reject unsigned cards? What isolates one tenant’s agents from another’s? Those are the questions that determine whether an A2A deployment is real or decorative, and they still land on the buyer.

Bottom line

The A2A move into AAIF is good news for anyone who worried about a single vendor steering the protocol, and it is less of a change than the headlines suggest. Google stopped controlling A2A in mid-2025. What the August announcement actually does is give MCP and A2A a shared venue to coordinate, which is genuinely useful.

What it does not do is standardize the hard parts. Enterprise authorization stays implementer-specific by the spec’s own language. Key-to-org binding, delegated authority, tenant isolation, and audit trails are not solved by either protocol. And conformance depth across the 150-plus organizations is unmeasured. If you are evaluating an A2A implementation, the consolidation changes nothing about the questions you need to ask. It just makes it a little easier to see who is actually answering them.

Sources

  1. A2A joins AAIF’s open agentic stack — AAIF
  2. Linux Foundation Announces the Formation of the Agentic AI Foundation — Linux Foundation
  3. Google Cloud donates A2A to Linux Foundation — Google Cloud
  4. A2A Protocol — official docs
  5. Agent2Agent Joins The Agentic AI Foundation, Alongside MCP — Forbes (as archived 2026-08-23)
  6. Enterprise Features — A2A Protocol