Agent fleet review — 1–7 Sep 2026
Produced by: Run EXAMPLE-FR-01 · read-only · document artifact
- Scope
- The week the review covers and the week it compares against, how many runs it read, and the rule it protected small groups with: a worker with fewer than three runs is withheld rather than reported.
- The fleet this week
- One row per worker — 6 named here — with runs, accepted, reworked, recorded failures, cost, median latency and trust. 214 runs, 167 accepted, $151.64 in total. Built-in agents, workspace workflows and outside runtimes sit in the same table, because they were all recorded the same way.
- What moved since last week
- Acceptance, rework and cost read twice — this week and the week before — so the page reports movement rather than a snapshot. Each line carries the runs it was read from.
- Failure evidence
- Recorded causes only, with how many runs each one covers. 4 runs had no cause recorded and are listed as "Cause not recorded" — never as a cause guessed from cost, latency or which worker ran.
- Cost
- Cost per accepted run, this week against last: $0.91 against $0.78. Total cost divided by accepted runs, and the page says so rather than leaving the reader to guess what the figure means.
- The three things to change
- 3 changes, each naming the worker, what to change and the evidence behind it — and each already filed as a ticket on the board, linked by its short id. Running the review again next week updates those tickets instead of filing them twice.
- Memory proposals
- Anything durable the review noticed, written as a cited candidate that stays unreviewed until a person approves it. The page never says memory was updated.
- What this review could not see
- The honest half: 4 workers withheld for having too few runs, 4 runs with no recorded cause, 2 workers with too little history for a trust rate, and a note that this review's own run appears in next week's roster.