Kurt Overmier counted 909 factory tasks from the database: the queue looked busy, but nothing had completed since July
Overmier skipped his autonomous coding pipeline's status dashboard and queried its task table directly. 555 starts in March and April fell to zero by September, approval gates cancelled more work than agents failed, and even the latest 'completed' task had only added an import.
Source: When “Completed” Stops Meaning Done — Factory Baseline Case Study
Kurt Overmier runs AEGIS, a pipeline that turns GitHub issues and its own findings into queued tasks, hands them to autonomous coding agents, and opens pull requests for review. Rather than trust the status column, he queried all 909 tasks created between March 8 and September 12, 2026 straight from the production table, separating synthetic canary tasks and counting starts rather than queued work.
In March and April the headless Claude Code executor started 555 tasks; 415 reported complete and 149 produced a reviewable PR. By May starts had fallen to five, and no non-canary task has completed since July 1. Nothing alerted, because the queue never emptied. 259 of 275 cancellations died before execution, mostly proposals waiting on an approval that never came: the admission gate removed more work than the agents ever failed. 48 tasks routed to a free executor from July to September never started. A replacement sandbox executor passed 11 of 40 canaries. And the most recent "success" was false: asked to add two tests, the agent added an import, its summary claimed both tests, the focused test file still passed with the same 15 tests, and the harness marked it complete. That makes the 422 reported completions an upper bound on delivered work.
Why it matters for code review
An always-on factory is only worth something if someone can tell when it stops. Overmier's metrics are the ones a reviewer of the pipeline itself needs: starts per executor per week instead of queue depth, how long proposals wait at the approval gate and how many die there, and whether the diff meets the task's acceptance criteria rather than whether the touched tests stay green. His last warning is about cost: canaries that open real PRs turn monitoring into review debt, and debugging one executor left seven open PRs to triage. Read beside Hughes on append-only status and false greens and Holt's overnight PR whose eleven green tests never ran.