A routine that runs is not a job that got done
Aug 2026 Paul Kuo: Cherny’s 388→180 maintenance PRs vs 150 days of launchd/Cowork jobs — silent failure when green logs lie; verification must sit outside the executor.
Source: The AI That Patrols My Systems Every Day — Until the Watchdog Itself Broke
Paul Kuo’s Aug 2026 essay is the productive complication for daily agent maintenance: a routine that runs is not the same as a job that got done. He reads Boris Cherny’s Anthropic experiment — Slack channel proj-claude-maintains-apps, daily Claude routines across iOS/Android/desktop/web/CLI/Agent SDK (crash fuzzer, dup unifier, dead-code remover with next-day log proof, flaky/useless-test cleanup, abstraction police), 388 PRs opened / 180 merged after Claude Code Review + human review — as a shift in who initiates work. The durable asset is operational knowledge encoded as routines you tune for tomorrow, not a single peak chat.
His own side is 150 days of defensive automation: 23 launchd jobs, crontab, hooks, and Cowork schedules (11 of 39 still active). The failure modes that matter for always-on factories all look healthy: exit green, logs full, dashboard lit — and nothing intended happened. Three varieties: (1) action ran with no effect (git add on paths later .gitignore’d, error swallowed by 2>/dev/null, ~90 days silent); (2) real alerts that were false (laptop sleep gaps); (3) the watchdog itself broken while reporting all clear (Aug 9 pre-push “health check” crashed, then printed hardcoded “no divergence” while origin was three commits behind, including a security fix). He changed the string to “divergence state unknown (detection did not complete)” — wording that must never revert.
The load-bearing design claim: verification must sit outside the executor. Cherny’s PR + automated review + human review + CI is an external auditable interface; the 388→180 gap is the funnel working, not a pure quality insult. Offensive routines fail loudly (no PR today); defensive ones produce nothing when healthy — indistinguishable from never having run — so roughly half his jobs exist only to prove others aren’t lying (contract curls vs hardcoded expected statuses, Sunday restore drills, chaos drills that manufacture failures). Routines also expire; disabled schedules keep a one-line “deactivated on / replaced by” note so six months later nobody fears turning them off.
Why it matters for code review
Continuous overnight maintenance without an independent “did it actually happen?” layer is FOMO cosplay. Green CI and a confident summary are the same costume Kuo’s watchdog wore. Always-on factories compound only when merge gates, contract tests, and restore drills refuse to let the agent grade its own shift — and when someone owns routine lifecycle so yesterday’s cron does not become unkillable folklore.