The Layer That Didn't Ship

The layer everybody agreed was missing

By late 2025 the diagnosis was common property. Agents were going to plan, negotiate and buy at machine speed, and they had no shared way to hand each other meaning. OpenRTB moves the request. AdCOM describes the ad. GPP carries the consent string. None of them answer the question one agent most needs to put to another: what is this moment about, and is this person worth reaching inside it?

LiveRamp proposed an answer. The User Context Protocol was a convention for passing embeddings between agents instead of shipping segment files around and hoping both sides had the same taxonomy loaded. LiveRamp donated it to IAB Tech Lab in late 2025. Tech Lab renamed it Agentic Audiences by early 2026 and filed it under the AAMP umbrella. The map I published in March 2026 called that layer the missing link and built half its argument on it.

In September I went to see what had been built. The answer is not the one the press releases imply, and the reason it turned out that way is the interesting part.

Eleven commits

IABTechLab/agentic-audiences was created on October 8, 2025 and last pushed to on July 27, 2026. In those ten months it took eleven commits. It has no releases and no tags. Two pull requests are open, against a repository whose last push was July 27, 2026. Its embedding-exchange spec is headed “Draft v0.1.” The repository still describes itself as “An open standard by LiveRamp defining how intelligent agents exchange user context, enabling interoperability from prompt-based LLM coordination to embedding-based optimization.” The wire format string in the repo is still vnd.ucp.embedding+json, which is to say the retired name is sitting inside the media type that implementers would have to code against.

Two things that paragraph does not say. It characterises a repository. It does not characterise a company’s intent, and I could find no LiveRamp statement on the work after April 2026, so I am not going to read intent into silence. It also does not say the effort is dead. A lightweight Commit Group was reported at the time as the governance for this work, and I could find no public charter, roster or meeting record for it. Absence of a public record is not evidence that nothing happened in private. Quiet on the public record since that last push is what the evidence supports.

The four weeks that shipped

Something real did get built, and it is worth being exact about how much. The OpenRTB community extension that carries these embeddings, sponsored by LiveRamp and Raptive, rides BidRequest.user.data with an aa envelope at Segment.ext.aa. It has nine commits, every one of them between March 22 and April 13, 2026. The Prebid.js real-time data module merged on April 14, 2026. The repository’s own Prebid folder calls itself a vendor example and points implementers at Prebid.js for the maintained version.

That is a working integration on two of the most widely deployed pieces of plumbing in the industry, delivered inside a four-week window, and nothing has been added to it since. What Agentic Audiences retains today is narrower than the map promised and still real: a payload convention that rides rails somebody else maintains.

The job got done twice, somewhere else

An interest to declare before the next two paragraphs: I co-lead AdCP’s Signals & Measurement working group and run a signals agent in its registry.

AdCP’s serve-time profile, Trusted Match, carries contextual embeddings in context_signals as base64-encoded int8 vectors alongside embedding_model and embedding_dims. It splits context matching from identity matching and separates the two in time, which is designed to stop them being rejoined; the spec states the requirement, and how well any given deployment meets it is a separate question. The clock is written into the spec: each provider gets a timeout_ms defaulting to 50 that has to fit inside the router’s latency_budget_ms, and an agent that misses the budget is dropped from the merged response rather than waited for. The surface is marked Experimental and its shapes can change on six weeks’ notice.

On the other side of the fence, AAMP’s in-bidstream runtime is ARTF, which reached v1.0 on July 20, 2026 and was pushed to on September 4. It runs OCI containers inside the host platform’s own cluster, speaks gRPC over vendored OpenRTB 2.6 protobuf, and returns atomic bid mutations the host accepts or rejects against a tmax budget.

Neither of those two specs references the other’s embedding vocabulary. The shared context layer the industry asked for in 2025 exists twice, in rival camps, with different field shapes and different rules about what may be carried. That is close to the opposite of what a donation to a neutral standards body was supposed to produce, and it is the clearest single illustration of the split I mapped in detail in July.

Why execution migrated to the runtimes

A neutral layer donated to a standards body is not the same thing as a layer that ships, and the gap between them is not a governance failure. It is a physics problem with a commercial shadow.

An embedding is only worth carrying if it arrives before the auction closes. Every design decision that matters about it follows from that constraint: how many dimensions, what quantisation, whether you can afford a network hop to fetch it, what happens to the response when the provider is late. You cannot settle any of those questions in a document that sits outside a runtime, because the answers are properties of a specific latency budget owned by a specific piece of infrastructure. Whoever holds the clock ends up holding the vocabulary, because they are the only party who can tell you whether a proposed field costs three milliseconds or thirty.

There is an incentive layer underneath the physics. The team that will be blamed when the bid response misses tmax is the team that runs the bidder. That team will not adopt a format it cannot change, and it will happily write its own. So the context layer landed where the clocks already were, in AdCP’s serve-time router and in ARTF’s in-cluster containers, and the donated convention stayed where it was filed, on a payload field of rails maintained by somebody else.

This generalises past ad tech. A standards effort with no runtime attached is a proposal for other people’s engineering time, and it will be executed on the schedule of whoever has that time free. A standards effort that ships inside a system with a performance budget gets maintained because the maintainers need it to work on Monday.

What to watch, on any standards effort

The tell is not the announcement, the umbrella it gets filed under or the governance structure named on the day of the donation. It is commits, tagged releases, and whether the one integration that exists is still being maintained. Those three take about ten minutes to check and they disagree with the press release often enough to be worth the ten minutes.

Run the same test on the camp I sit in and it gives a mixed reading rather than a flattering one. AdCP tagged v3.1.21 on September 9 with v3.2.0-rc.1 in prerelease, so the release cadence is real. Its registry holds 35 agents. But a discovery sweep on September 10 found only 9 of them answering a public tools/list: 8 sit behind authentication, 13 return an error, 5 are unreachable. In July the equivalent sweep found 7 discoverable agents out of 21. The roster grew by fourteen and the number of agents a buyer can actually discover grew by two. Read that as commits and releases running ahead of deployed, reachable code, which is a different failure from the one Agentic Audiences has but not a reason to be smug about it.

The map I wrote in March treated one shared context layer as the keystone of the agentic stack. It was the right thing to look for and the wrong place to expect it. Context arrived as a feature of two competing runtimes because that is where an embedding has to land on time to be worth sending at all.

Here is the checkable version. If Agentic Audiences is being maintained rather than parked, the first visible sign will be cheap and specific: a tagged release, or a commit retiring vnd.ucp.embedding+json for a media type carrying the name the spec has had since early 2026. If neither appears by the end of 2026, and neither camp’s serve-time spec has added a reference to the other’s embedding fields by then, the context layer is settled and the settlement is two vocabularies, permanently.

Addendum, September 16, 2026 — two things in this essay need adding to, and the second is a miss of mine.

The repository moved, slightly. On September 15 a collaborator was added to IABTechLab/agentic-audiences, and half an hour later that collaborator closed pull request #17, a documentation link to the OpenRTB extension, noting that the link was already on the main branch. That leaves one open pull request, #11, with no activity since April 13. Closing #17 was the first maintainer action on any pull request or issue since the July 27 push, and it reads as tidying rather than shipping: there is still no commit since July 27, no release and no tag, and the wire format string still reads vnd.ucp.embedding+json. Neither of the two signs this essay closes on has appeared.

The governance point needs adding to. I wrote that I could find no public charter, roster or meeting record for the Commit Group reported as this work’s governance, and that still holds for the Commit Group. But on August 4, five weeks before this essay, IAB Tech Lab’s chief executive Anthony Katsur published a post saying the work to date had been driven by the Tech Lab Agentic Task Force, and announcing a separate Agentic Audiences Task Force with sign-up through Tech Lab’s membership team. I should have found it. It also means the public record was not quiet after the July 27 push, as I wrote; the repository was. The same post credits an OpenRTB community extension and a Prebid module update with making the data exchange “interoperable, but also transaction-ready”, which is the integration this essay identifies as the part that shipped. The post also names an open-source scorer, which this essay did not discuss; it lives in the same repository, and the July 27 commit was a fix to it.