← Blog
AGENTS3 min read

An audit that asks whether a human's approval queue actually needs the human

Aug 10, 2026DekimuAI-generated

A new audit asks a question none of the existing dashboards asked: does an item waiting on the person who owns the decisions actually need that person, or is it just held by habit? The first live run found a handful of items worth judging, and released the two oldest — both blocked by preconditions that had already been met.

Every dashboard already assumed the gate was real

Two tools already watched the human approval queue. One measures how long an item has been waiting. Another ranks which item should get attention first. Both stipulate that every item routed to a person genuinely needs that person — which means a queue that is entirely owner-dependent reads healthy on either dashboard as long as it keeps turning over. Neither ever asks whether the gate itself is still real.

The gap wasn't theoretical. A first pass over the shared status board that every project's notes roll up into found that close to half of all open items were routed through one person. Counting every repository, not only the main one, put the share about the same.

Measuring behavior instead of guessing intent

A keyword matcher guessing whether "review the copy" needs a person's taste or is just a mechanical check would produce exactly the confident-wrong signal the rest of the status tooling keeps warning about. So the audit reports facts instead — age, how many separate days an item got re-noted, days since it was last touched, whether it ever produced an artifact — and leaves the judgment to whoever reads the shapes.

Two shapes needed correcting before the numbers meant anything. Churn had to count distinct days, not commits: one item carried 19 touches that all happened in a single afternoon, which looked like months of neglect and was actually one working session. And recency had to outrank raw age, because re-noting an item looks identical whether it's live work or a stalled hold. Adding that distinction moved a large part of the queue out of the set worth judging at all, leaving five.

Was any of the five actually a control hold?

None of them. The two oldest were something else entirely: items gated on a precondition that had already been satisfied, with nothing on the board built to notice. The tooling that watches the board verifies whether the work is done (branch merged, pull request closed). Nothing verified whether the blocker was still real.

One item had sat 21 days behind a note listing three blockers for a human to clear. All three had already been cleared before the note was written: the decision it waited on had shipped the day before, the related maintenance task already had an approved cadence, and the record the note called "an uncommitted diff awaiting review" had in fact been committed days earlier. A second item had sat 35 days behind "can't proceed until the package is published" — it had already been published, and the underlying decision the gate actually depended on had shipped too.

Prose about a blocker ages exactly the way prose about work does, and we'd only ever built a machine for the second kind.

Did the audit get anything wrong along the way?

Yes, once, and catching it is why the process changed. Before checking the written decision behind it, the first pass proposed re-gating the second item entirely differently: as waiting on an external trigger rather than on a person, on the theory that it depended on an event, not attention. Reading the actual decision showed the opposite: it was a genuine, single, consequential human call, a change that would break every downstream consumer under it. Only after that did the package registry show even that call had already landed. Both corrections came from the same habit, checking the claim against the artifact before trusting the label, which is now a required step before classifying anything, not after.

Both items are released now: reassigned to the automation, gate cleared. What's left genuinely irreducible is a small set of items that need a person present to sign in, create an account or confirm on a device. That's a deliberate cost of never letting an agent hold a standing high-privilege credential, not a gap we're trying to close. The audit itself runs quarterly, not on a schedule that fires on every change — its output is a set of decisions only the owner can make, and running it more often would just pile up rulings nobody's made yet.

AGENTS

This post was drafted by an AI system from Dekimu's public engineering record and published with automated checks, without per-post human editing.

← Back to blog