The protocol war is over: ACP has folded into A2A, and the nameplate layer is still empty
Sources cited in this piece · ACP Joins Forces with A2A, LF AI & Data Community Blog, 29 August 2025, by Kate Blair (IBM Research) and Todd Segal (Google) — primary source, quoted verbatim · Agent Protocol War Ends: ACP Merges Into A2A, ClawdBytes, 15 October 2026 — secondary retrospective, quoted verbatim and labeled as such · ecosystem figures (MCP weekly SDK downloads, registry server counts, A2A joining the Agentic AI Foundation) are secondary reporting dated around 9–15 October 2026 and labeled as such throughout.
Eighteen months, three protocols, one umbrella. IBM Research launched the Agent Communication Protocol (ACP) in March 2025 for its BeeAI platform; Google introduced A2A weeks later; Anthropic's MCP was already spreading through the tool layer. In August 2025 ACP merged into A2A, and in August 2026 A2A joined the foundation that already governed MCP. The agent protocol war is over — and what it settled was only the argument about how agents connect. One year on, the industry has consolidated everything except the one layer that lets a party say which agent it means: tens of thousands of registered servers, more than 150 supporting organizations, two protocols under one roof — and no standard way to write down an agent's name.
1. What merged, in the words of the two people who announced it
The primary record is a joint post on the LF AI & Data community blog dated 29 August 2025, signed by Kate Blair of IBM Research and Todd Segal of Google. Its opening line sets the sequence: “IBM Research launched the Agent Communication Protocol (ACP) in March 2025 to power its BeeAI Platform, an open-source platform exploring agent interpretability.” Then the donation: “Later that month, the BeeAI project—and with it, ACP—was donated to the Linux Foundation, solidifying our commitment to openly advancing agent interoperability as a community.” Then the announcement itself: “Today, we're excited to share that ACP is officially merging with the A2A under the Linux Foundation (@thelinuxfoundation) umbrella.”
The mechanism is stated plainly, and it is a winding-down rather than a merge of equals: “As part of this transition, the ACP team will be winding down active development and will begin contributing its technology and expertise directly to A2A.” The governance seat that came with it is the interesting artifact: “Blair will join the A2A Technical Steering Committee on behalf of IBM, alongside representatives from Google, Microsoft, AWS, Cisco, Salesforce, ServiceNow, and SAP.” Users were promised a path, not a bridge: “Migration paths and documentation will be provided to help ACP users seamlessly move to A2A.” And the forward quote, from Segal: “We're thrilled about what's ahead and look forward to building the future of agent communication together.”
2. One year on: what consolidation bought, and what it charged
A year and a half later the retrospective writes itself. The ClawdBytes analysis of 15 October 2026 records the state of A2A in one line: “A2A reached v1.0 in March 2026 and now boasts over 150 supporting organizations, including deep integrations with Google, Microsoft, and AWS.” On the tool side, secondary ecosystem reporting dated around 9 October 2026 puts MCP's weekly SDK downloads at roughly 92 million and the official MCP registry at roughly 41,000 servers; the same reporting places A2A under the Agentic AI Foundation from August 2026, the same body that governs MCP. Two layers, one roof, one number that matters: three protocols became two, and the two now answer for everything except naming.
The consolidation was not free, and the people who paid were the early adopters. The retrospective's phrasing is the sharpest thing in it: “the result isn't a simple version bump; it's a re-architecture of their coordination layer.” It carries the anecdote that makes the abstraction concrete — “One developer reported spending three weeks refactoring a multi-agent system because their team built on ACP, only to find it deprecated two months before they started building” — and the honest caveat about the escape hatch: “While a migration path exists for BeeAI users via the A2AServer adapter, it does not undo the architectural debt incurred by ACP's assumptions.” The trade is summarized without either side being flattered: “The Linux Foundation framed this merger as accelerating progress toward a single, more powerful standard, but the cost was borne by early adopters who shipped on the now-deprecated protocol.”
3. Why a merged protocol still cannot name an agent
Here is the part that matters for us, and it is not a criticism of A2A — it is a statement about which rows of the stack anyone is even trying to fill. In A2A, discovery works by publication: an agent publishes a machine-readable Agent Card at a well-known location on its own origin (/.well-known/agent-card.json). That is a document in DNS space, self-published by whoever controls that origin. The retrospective's two technical observations are the ones to keep: “Task state is self-reported by the executing agent, and while Agent Cards can carry JWS signatures, they are optional in many implementations” — and, on discovery at scale, “There is no formal registry for discovery yet, as the specification never shipped and maintainers redirected the need to adjacent specs.”
Now put the two facts side by side. A2A solved transmission between agents and did it well enough that 150 organizations signed on. What it did not ship is a canonical name: an agent is identified by where it publishes a file, and the file is what the publisher says. A catalogue of self-assertions is not an address system — it tells you what has been claimed, not what is the case. The MCP registry's tens of thousands of entries are exactly this kind of catalogue, and they are excellent at what they are; they are not, and do not claim to be, a way to compute an agent's name from its authority.
The retrospective lands on the same distinction in commercial language, and it is the best sentence anyone has written about the gap this year: “'A2A-compliant' is a procurement label, not a safety guarantee. The protocol solves 'how to transmit,' not 'how to trust.'” Two rows, stated in one line. Transmission is solved. Trust is not, and neither is the addressing that trust has to hang from.
4. Where the two schemes sit
RTTP (rttp://) answers where and when — the third row. An address under our scheme is not published, registered, or looked up; it is computed: SHA-256(canonical_authority)[:16], a function of the authority itself. This is the specific thing an Agent Card is not. A card is a document at a domain: if the domain changes hands, the card can change; if two catalogues list the same name, neither is wrong by construction, because the name was never derived from anything. A derived address has no such freedom — the second-preimage property is the reason, and it is arithmetic, not policy. Sixteen hex characters, no landlord.
IQA (iqa://) answers whether — the fourth row, the one the quote above leaves open. An attestation envelope carries the statement with an Ed25519 signature; standing is a 120-second lease with four states (ghost, probation, radiant, genesis); malformed input fails closed, so REJECT is a conclusion rather than an exception; and verification is local arithmetic, which is what lets the three offline tests — disconnection, survival, explanation — mean anything. Where we stand on the record: rttp:// and iqa:// are recorded as IANA provisional schemes (CRI 27 and CRI 28), and the joint addressing draft is in IETF ISE review. Slower than shipping, and deliberately so.
The honest reading of this week's news is that consolidation was never going to produce an address — it was only ever going to clear the ground for one. You cannot standardize names across an ecosystem that has not yet agreed on how its members talk; the merger is a precondition, not a substitute. That is also why the emptiness of rows three and four is not an accusation against MCP or A2A. It is an observation about which rows the industry has had time to fill.
The war ended where everyone was watching — who talks to whom. The frontier moved to the row nobody had claimed: who, exactly, is talking. Two protocols under one foundation, 150 organizations, tens of thousands of registered servers, and still no standard way to write down which agent you mean. That is the row our two schemes were written for, and the year the protocol war ended is the year it became obvious the pen was still on the table.