Developer builds self-routing AI agent fleet that files, tracks, and reports its own work

A software developer has redesigned their multi-agent task system so that AI agents now detect, file, and route their own work each night without human intervention. A curated allowlist determines which finding classes — such as stale paths, dependency advisories, and outdated CI tests — can be automatically queued, while judgment-heavy tasks are reserved for human review. Progress is now measured by whether a queue entry actually advances, not by the number of commits made, preventing agents from appearing productive while leaving real tasks untouched. Entries that fail to progress over three separate nights are automatically flagged for the owner with a note. The accompanying morning briefing was redesigned to a concise, phone-readable summary under one minute, condensing a 4.6 KB nightly report down to 680 characters covering only what requires human attention.