The conformance census measured what eight agent systems can record. Of the six that take or gate actions, one can name the person who approved one. That is a capability gap, and a capability gap is not the same as a need. A system can lack something nobody ever wanted.

So this is the other half, and it is the half more easily faked. It asks who has requested this capability, in public, where the request and its outcome can both be read.

The distinction that decides the answer

A naive count is worthless here, because two different things are happening in these issue trackers and they look identical from a distance.

• Somebody with the problem opens an issue on the framework they use, once, describing what they cannot do.
• Somebody selling a solution opens issues on several frameworks, often naming their product, describing what their layer provides.

The second is evidence about suppliers. Only the first is evidence about demand, and counting them together produces a number that says the market is asking when what it is really doing is selling.

The rule used here is mechanical, so it can be argued with: an issue is promoting if its author filed on two or more of these repositories, or if its title names a governance product. Otherwise it is asking. It is a proxy for intent and it will be wrong in individual cases. The data is published so that any individual case can be checked.

What the split looks like

Eighty-four issues with the subject in the title, across seven agent frameworks, from fifty-one distinct authors. Most of the traffic is supply, not demand. Fifty-two of the eighty-four come from people with something to sell, and thirteen authors account for the bulk of them by filing the same case on several frameworks at once.

The genuine requests are outnumbered and better attended. Each one draws three times the discussion of a promotional issue, which is what you would expect if the people replying can tell the difference.

The finding

Thirty-two people asked for this capability without having one to sell. Twenty-two of them got no resolution. Fourteen of those issues are still open and eight were closed as not planned.

The most discussed of those thirty-two, an RFC for tamper-evident audit trails carrying ninety-eight comments, was closed as not planned. The two after it, at sixty and forty-six comments, are still open: a cryptographic governance layer, and guidance on auditable completion receipts.

The two busiest threads in the whole set are larger still, at three hundred and eighty-three comments and a hundred and twenty-nine, and both are counted here as promoting. They are worth reading anyway. A vendor thread that draws that much argument is not evidence that nobody wants the thing.

Set against the census, the two readings say the same thing from opposite ends. The systems cannot record who approved an action. People have asked them to. Most of those requests are unanswered.

The frameworks are not planning to close it

This is the part that makes the gap structural rather than a backlog. Closing an issue about compliance hooks, a maintainer of one of these projects wrote:

“we don’t plan to have some examples within this repo. Please feel free to share examples in your repos or docs.”

That is a reasonable position for a framework to take. Orchestration libraries are not authorisation layers and never claimed to be. But it means the capability is being pushed outward rather than scheduled, and a deployer waiting for their framework to grow it is waiting for something nobody upstream has undertaken to build.

What this does not show

It does not measure demand. It measures public requests on seven repositories, which is a small and unrepresentative window on it. A search finds what it is spelled to find. Anybody who needed this and solved it privately, asked in a support channel, or never asked at all is invisible here, and that is most people.

Nor does an unanswered issue prove a maintainer was wrong. Projects decline things for good reasons every day, and several of these were declined with one.

The claim is narrower than eighty-four looks: most of the traffic is vendors, a real minority is not, and the majority of that minority went unresolved. Thirty-two is the number that means anything here, and twenty-two is the one worth arguing about.

Check it

Every issue, its author, its state, the number of repositories that author filed on, and the classification this reading gave it are in census/demand/issues.json. The counts on this page are recomputed from that file by the test suite rather than typed, so a disagreement is settled by reading the data.

If a classification is wrong, it is wrong in a way a pull request can fix, which is the same standing every verdict in the register has.

Citing this

The readings are CC BY 4.0, which asks for attribution, so the reference is here rather than left to be composed. This block is generated from the page it sits on, so a date that moves here moves in the citation too.

Clifford, T. (2026). Who asked agent frameworks for audit trails and approval records. Machine Testimony. https://machinetestimony.org/demand/

@misc{clifford2026demand,
  author = {Clifford, Troy},
  title  = {Who asked agent frameworks for audit trails and approval records},
  year   = {2026},
  note   = {Machine Testimony, read 6 September 2026},
  url    = {https://machinetestimony.org/demand/},
}

This page carries no DOI. It cites its dated URL, and saying so is the point: a citation naming a deposit that does not exist is worse than one naming a page that does.