A2A and MCP now share a home. What the Agentic AI Foundation settles, and what it leaves open

Published October 23, 2026 · Last reviewed October 23, 2026

On 17 August 2026 the Agentic AI Foundation announced that Agent2Agent (A2A), the protocol agents use to find each other and hand over tasks, had joined it as a hosted project.¹ The foundation already stewarded the Model Context Protocol (MCP), the protocol agents use to reach tools and data. Both were already vendor-neutral Linux Foundation projects. Now the two protocols most agent builders mention in the same breath share one foundation.

In governance terms the move is a consolidation. In technical terms almost nothing changed. A2A is still v1.0.1, released on 28 May 2026.² The two projects keep separate repositories and separate technical leadership. And the hardest open question in this part of the stack, how a stranger’s agent reaches a specific business or person who has agreed to be reached, is on neither project’s agenda.

AGNTCon + MCPCon North America, the foundation’s conference in San Jose, opened on 22 October. Day one brought four new AAIF projects and a member count of 305, but no dates for MCP’s identity work, no new A2A release and no discovery effort taken on by the foundation.³⁴

How A2A got into the Agentic AI Foundation

The Agentic AI Foundation (AAIF) is a vendor-neutral directed fund under the Linux Foundation. It was announced on 9 December 2025, co-founded by Anthropic, Block and OpenAI, with MCP, goose and AGENTS.md as its founding projects.⁵ It is not Anthropic’s foundation. AAIF’s own A2A announcement maps its projects to layers of the stack: AGENTS.md for instructions and context, goose as an agent runtime, MCP for agent-to-tool connectivity, agentgateway for traffic mediation, and A2A for agent-to-agent interoperability.¹ On 22 October it added Agent Skills (contributed by Anthropic), Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba), for 10 hosted projects in all.³

A2A took a short, documented path in. The proposal was filed in AAIF’s public proposals repository on 18 June 2026, with a contact from Google’s open source programs office. AAIF’s Technical Committee approved A2A as a Growth Stage project in July, and the Governing Board approved inclusion on 4 August, when the proposal was closed as complete. Because A2A was already a Linux Foundation project, a comment in the thread notes that “no further paperwork is needed.”⁶ The announcement followed on 17 August, and the A2A project published its own post on 27 August.⁷

Membership has grown fast. AAIF reported 170 member organisations in April 2026.⁸ On 13 August, four days before the A2A announcement, the Linux Foundation put the total at 247, with Alibaba, Visa and Wells Fargo joining at Gold level.⁹ On 22 October, in San Jose, it said 305.⁴ AAIF also says A2A has more than 150 supporting organisations, some of them direct competitors.¹

What one neutral home settles

The first thing the move settles is who holds the keys. A2A was created at Google, launched in April 2025 and moved to the Linux Foundation that June; Anthropic open-sourced MCP in November 2024, and MCP moved to AAIF at its launch.¹⁰ So both were already vendor-neutral. What changes is that a company building on both now deals with one steward, one membership and one set of project rules instead of two, and A2A’s proposal committed the project to hand its trademarks and accounts to AAIF on acceptance.⁶

The second is a shared venue. Agent-to-tool and agent-to-agent work now meet at the same conferences, under the same membership and the same lifecycle rules. The A2A project’s post says “foundational infrastructure must evolve openly and predictably.”⁷ That matters most where the two protocols touch: an agent that calls tools over MCP and delegates work to another agent over A2A now has one governance body to watch rather than two.

Just as clear is what the move does not merge. A2A and MCP stay separate projects with their own technical leadership, and AAIF still points A2A contributors to the project’s own steering committee meetings.¹ There is no joint protocol specification. The closest thing is AI Catalog, a draft format for listing a domain’s MCP servers and A2A agent cards in one file. It lives in a temporary Linux Foundation repository, it is not an AAIF project, and the A2A and MCP steering committees are to vote on adopting it once it is final.¹¹ There is no new A2A version either: builders have nothing to migrate, and the Agent Card still lives at /.well-known/agent-card.json.

WebMCP, the browser interface that lets a page expose tools to an agent through document.modelContext, sits outside AAIF too. It is a Draft Community Group Report of the W3C Web Machine Learning Community Group, updated on 20 October 2026, and not a W3C Standard.¹² So the three pieces people often draw as one agent stack (MCP for tools, A2A for agents, WebMCP for the browser) still have two homes. The stack is an emerging mental model, not a jointly designed standard, as our three-layer stack explainer sets out.

What the MCP roadmap puts on the identity agenda

Five days after the A2A announcement, on 22 August, MCP maintainers David Soria Parra and Den Delimarsky published “The New MCP Roadmap”. It names five priority areas: agentic messaging primitives, transport unification and hardening, agent identity and enterprise-ready security, improved primitives, and a better SDK developer experience.¹³

The third area is the one that touches identity. Its goal is a standard way to recognise and trust agent identities. The roadmap plans to finish Demonstrating Proof of Possession (DPoP) and drive its adoption, and to define “an opinionated path for agent identity and delegation” using workload identity federation, the ID-JAG grant and standard token exchange, including handing narrower authority to sub-agents. It also promises more work with the IETF’s OAuth and WIMSE groups.¹³ The authorization side is covered in our piece on MCP enterprise-managed authorization.

This is identity for the machine side: which workload is calling, on whose delegated authority, with a key it can prove it holds, and with what narrower permissions for a sub-agent. Enterprises have that problem now, and vendors are shipping against it. On 22 September Okta made Agent SSO and Agent-to-Agent Connections generally available; the second controls which agents may call which and records each handoff.¹⁴ On 22 October, at the San Jose event, Google Cloud described a SPIFFE-based identity for each agent and a registry of approved agents and MCP servers in its own governance tools.¹⁵ These are products, not AAIF standards.

The roadmap has three limits. It gives no dates, only the coming months and the next specification release, and on 23 October the DPoP proposal (SEP-1932) was still an open draft.¹⁶ A pull request opened that day would list the DPoP profile as an experimental MCP authorization extension, a stage with no stability promise; it gives no dates.¹⁷ Its discovery work is progressive, not public: a server offers a small entry point and reveals more tools as a conversation narrows, which says nothing about finding servers or agents on the open web.¹³ And Server Cards sit outside the five areas, though they moved anyway: SEP-2127, which defines MCP Server Cards as an optional extension, is now marked Final, and it leaves domain-level discovery to an AI Catalog.¹⁸

The identity roadmap belongs to MCP, not to A2A. A2A’s own roadmap, updated on 15 September, lists v1.1 refinements, bidirectional streaming, a command-line tool and elicitation, with no identity item.¹⁹

What it leaves open: finding a specific business, and consent

Put the August move and the MCP roadmap side by side and one gap stands out: both assume you already know which server or agent you want to talk to. Neither says how an agent acting for a buyer finds a specific accountancy firm, clinic or consultancy it has never met, or how it learns whether the person behind that business agreed to be contacted by agents at all.

The card itself is the closest A2A comes to answering the first question, and it is still rare on the open web. Kin Lane of API Evangelist counted 421 of the 27,840 providers in the APIs.io catalog serving an A2A agent card from their own host on 21 September 2026, about 1.5%. Only 4,718 providers had been probed, so he treats 421 as a floor.²⁰ A card, or a catalog that lists several, tells you what an agent does and how to reach it once you are at the right domain. It does not get you to the domain, as our explainer on how A2A agent cards work sets out.

Work on discovery is happening outside AAIF, at an early stage, and only one piece of it tries to get an agent to a domain it does not already know. At the IETF, DAWN held a working-group-forming session at IETF 126 in July, was renamed Discovery of Agents With Names in August, and became a chartered working group on 22 October, the morning San Jose opened. Its charter covers discovery on a local network, inside an organisation and between cooperating parties. It leaves out agent identity management and “AI resource indexing and searching across the wider Internet,” and its first protocol specifications are due at the IESG by the end of 2027.²¹ Agentic Resource Discovery (ARD), a v0.91 proposal whose authors include engineers at Google and Hugging Face, has a site publish one manifest at /.well-known/ard.json listing its MCP servers and agent cards, and defines a search API that registries can run across many such manifests.²² Sierra and Meta’s Personal Agent Protocol, published as Draft 0.1 on 9 October, has a company publish /.well-known/poppy.json to tell personal agents how to interact with it and how customers sign in. It is not about introductions between strangers.²³

The second question, consent, is further out still. Workload identity can tell a service that an agent is who it claims to be and acts on someone’s delegated authority. It says nothing about whether the human on the receiving side wanted the conversation. For an enterprise calling its own tools, that question barely arises. For a small business receiving agents on its website, or a professional deciding whether a stranger’s agent may pass on their contact details, it is the whole question.

A foundation that stewards transport and tool protocols is the wrong place to decide who should meet whom. But the governance news should not be read as a sign that agent identity is handled. The machine side is getting clear owners. The human side is still being built by others.

What San Jose announced on day one

AGNTCon + MCPCon North America opened at the San Jose McEnery Convention Center on 22 October, the first North American edition of the conference; the European one ran in Amsterdam on 17 and 18 September, a month after A2A joined.²⁴²⁵ A simple test helps read the announcements. Dates or a draft for MCP’s identity work would be news. So would joint A2A and MCP work on discovery, a discovery project that AAIF takes on formally, or a new A2A release. More members and more logos would not.

Checked against what AAIF, the Linux Foundation, the MCP and A2A projects and Google Cloud had published by 16:40 UTC on 23 October, day one scored like this:

What came instead was the category that does not count: four new projects and 305 members.³⁴ AAIF’s executive director, Mazin Gilbert, said the new members are asking “how agents authenticate, how their actions are verified” and who is accountable, which is the machine-side agenda described above.⁴ A2A had no keynote of its own, but the day-one program included a hands-on A2A workshop and a technical session on A2A v1.²⁴

Day two runs on 23 October, with a Google Cloud keynote on MCP, A2A and OpenTelemetry, an Anthropic keynote on building with MCP, and the two AI Catalog sessions.²⁴ By 16:40 UTC, AAIF had published one day-two post, on its first Technical Excellence Awards, with no identity, discovery or release news.²⁶ On the MCP side, the pull request that would make DPoP an experimental extension was still open. If later sessions change the table, we will update it here.

After San Jose the calendar moves elsewhere. IETF 127 in San Francisco (14 to 20 November) is the first meeting since DAWN and AGENTPROTO, a new group on agent communication protocols, were chartered, and Chrome’s WebMCP origin trial runs through Chrome 156, with a desktop extension to Chrome 162 filed on 28 September.²⁷²⁸¹² Neither is an AAIF decision, and both bear more directly on how agents find and work with a business than anything the foundation announced on day one.

How this connects to Tobira

Tobira builds on these protocols and owns none of them. The gap described above, finding a specific business or person and reaching them only with consent, is the part we work on, as a human-facing layer next to A2A, MCP and the discovery drafts, not in place of them.

One public example: Relovisa, an immigration consultancy that shares a founder with Tobira, serves an A2A agent card on its own domain at relovisa.co/.well-known/agent-card.json. As it read on 23 October 2026, the card carries both the older v0.3 fields and a v1.0 JSON-RPC interface, points to an endpoint hosted on Tobira, and says the agent shares contact details with the Relovisa team “only with the person’s explicit consent”.²⁹ On the website side, Tobira Site Agent answers and sells to people and their AI agents, so a visiting agent has someone to talk to once it arrives. For why that differs from publishing files alone, see agent-readable vs agent-addressable.

What to remember

• The move is governance, not a new spec. A2A was approved as an AAIF Growth Stage project on 4 August and announced on 17 August 2026. It is still v1.0.1, and builders have nothing to migrate.
• One steward, two projects. MCP and A2A share a vendor-neutral home under the Linux Foundation but keep separate specifications and technical leadership. AI Catalog, their shared discovery draft, sits outside AAIF, and WebMCP stays at the W3C.
• Identity work so far is MCP’s, and machine-side. The 22 August roadmap targets DPoP, workload identity federation and delegation to sub-agents. It has no dates, and the DPoP proposal is still a draft.
• San Jose brought projects and members, not dates. Day one added four AAIF projects and took membership to 305. Discovery of a specific business is still outside AAIF: the IETF’s new DAWN group leaves open-web search out of scope, and ARD and the Personal Agent Protocol are drafts.

IETF 127 opens in San Francisco on 14 November, the first meeting since DAWN was chartered.²⁷ That is where we pick up the discovery half of this story next.

Frequently asked questions

When did A2A join the Agentic AI Foundation?

AAIF’s Governing Board approved A2A as a Growth Stage project on 4 August 2026, and the foundation announced the move on 17 August 2026. The A2A project published its own post on 27 August. The proposal had been filed on 18 June 2026, with a contact from Google’s open source programs office.

Did the A2A specification change when it joined AAIF?

No. The current release is still v1.0.1 from 28 May 2026, and v1.0.0 shipped on 12 March 2026. Joining AAIF changed who stewards the project, not the protocol. The Agent Card is still published at /.well-known/agent-card.json. A2A’s roadmap lists v1.1 work, but no v1.1 had shipped by 23 October 2026.

What did AGNTCon + MCPCon North America announce?

On day one, 22 October 2026, AAIF announced four new hosted projects (Agent Skills, Open Responses, OpenEnv and OpenSandbox), taking it to 10, and said membership had reached 305. There was no new MCP specification, no new A2A release and no dated plan for MCP’s identity work. By 16:40 UTC on 23 October, day two had added an AAIF awards post and an open MCP pull request to make DPoP an experimental extension, but no dates, releases or discovery projects.

What does the MCP roadmap say about agent identity?

The roadmap of 22 August 2026 lists agent identity and enterprise-ready security as one of five priority areas. It plans to finish DPoP, define a path for agent identity and delegation using workload identity federation and token exchange, and work with the IETF OAuth and WIMSE groups. It gives no dates. On 23 October 2026 the DPoP proposal was still an open draft, and a pull request opened that day would list it as an experimental extension.

Does the Agentic AI Foundation solve agent discovery?

Not as an AAIF project. MCP Server Cards are now a final MCP extension, and AI Catalog, a draft format developed by MCP and A2A contributors outside AAIF, lets a domain list its MCP servers and agent cards in one file. Finding a specific business an agent has never met is still outside both. The IETF’s DAWN working group, chartered on 22 October 2026, leaves search across the wider internet out of scope, and ARD and the Personal Agent Protocol are still drafts.

Sources