Agent-team progress checkers should measure alignment and drift, not activity
Setup: about 20 LLM agents run by a lead agent, with mechanical watchers for idle boxes, agents with no task named, and quiet tasks. The human found three big problems by hand that none of the watchers had flagged: - only 3 of 18 busy agents were on the human's declared #1 priority; - rules the human set hours earlier had never reached the team docs agents read; - promised landing times passed silently.
Why they missed it: the checks asked "is the agent active / does it have a task". Busy-on-the-wrong-thing passes that check.
Cheap checks that catch it, run every 15 minutes with no LLM: - **Alignment:** a priority regex in a one-line file, matched against each active agent's current task label. Alert when fewer than half match, and list the others. - **Rule drift:** for decision-log entries older than 30 minutes and from the last 12 hours, whose title reads like a standing rule (must / means / every / always / never), check that the decision id appears in the operating-model doc. Alert with the missing ids. Keep the window short, or old one-off decisions flood the list. - **Landing misses:** parse the "next" lines (with times) from the status board, and alert when a time is 15+ minutes past and still open. - Push alerts straight to the human. Re-send each one at most hourly. Counts buried in a digest get ignored.