Everything we have written about Automation.
My agent fleet's task board had an hourly sweep that moved closed issues to Done, and 50 closed issues still weren't Done. One board read showed why: every stuck card sat in a status the sweep never checked. The fix widened one list, cut the sweep's API cost by 99%, and taught it which cards it must report instead of touching.
My agent fleet's CI started failing 3 to 18 seconds into every job. The cause was two GitHub Actions runner listeners sharing one work directory, left behind by a runner self-update and KillMode=process. The diagnosis, and the ExecStartPre fix.
My merge-card scanner caught assertions being removed but never noticed a new test that had none. Closing that gap meant giving up on syntax trees, teaching the scanner to know when it is looking at a cropped view of a file, and measuring the false-positive rate on thirty merged pull requests before the line ever reached a human.
I pressed Merge on four agent pull requests and only one merged. The bug was an ack sitting upstream of the action, a serial update queue I was silently relying on, and a merge digest that could not see two PRs colliding on the same files.
A three-layer quality gate for AI-generated articles: deterministic checks, a rubric-based LLM judge, brief-grounding verification, bounded revision, and final human approval. With real numbers from the gate's first run, which this very article failed at 5.0/10.