One cron job at 01:00. Twenty AI agents fan out across every repo and every system on the machine, then file a single ranked report — done before I pour the coffee.
A growing AI workspace drifts. Repos go stale, a secret can slip into a commit, follow-ups stall with no nudge — and you only notice when something breaks. The missing piece isn't more building. It's oversight: a standing check that runs whether or not I remember to run it.
One Workflow script does the fan-out: 15 per-repo audits (health + security on every git repo on the machine) plus 5 deep system scans (config inconsistencies, contacts, operational follow-ups, memory, and a 90-day git story). Windows Task Scheduler fires it via claude -p; a small lock file stops a second run from overlapping; the output lands in one dated summary on disk.
Three mornings of real reports in, the security posture came back clean — no critical or high findings across all 15 repos. Good news, and the least interesting half. The value was the drift it caught: a proposal quietly due in 48 hours, a high-leverage action that had sat untouched for five weeks, and five stale follow-ups a 15-minute sweep would clear. Things a human stops noticing; a nightly agent never does.
Twenty agents on a top model at 01:00 is not cheap — the internal name was earned. The trade is simple: a few cents of tokens a night buys an audit I would otherwise never sit down and do. agent_lock stops a double-run from doubling the bill.
Every run leaves evidence — the fan-out as it happens, the two phases, the green security table, and the ranked action list waiting at breakfast. Tap any image to enlarge it and read the exact prompt that drew it.
TL;DR bullets, a severity-sorted security table, and the five highest-leverage actions — generated by twenty agents at 01:00 and sitting in a dated file before I opened my eyes. Three mornings running, no clicks from me.
Why I let twenty agents loose on my own machine every night — and read the report before coffee.
A Workflow that fans out one agent per repo plus a few whole-system scans, fired by your scheduler at night, writing one ranked report to disk. Everything in this episode is free and open — clone it, run it, make it yours.
schtasks /create /tn overnight-audit /tr "claude -p -f overnight-wf.js" /sc daily /st 01:00
No repo yet — comment AUDIT on the post and the bot DMs you the workflow script the moment it's public.
Next: the agent that doesn't just report the drift — it opens a branch, drafts the fix, and leaves it waiting for a yes or no. From night-watchman to night-shift worker.