Topic
Automated workloads, like scheduled jobs or AI agent runs, that execute without a human present to approve actions or notice failures.
An unattended system needed five interventions in two days. None of them printed an error a human would see, because every one left the process running and the watchdog green. The fixes are ordering rules.
Reality
Seven Claude Code subagents produced 382 commits, 24 board meetings and 16 articles in 21 days, and no revenue. The reusable part of the writeup is the runner that treated exit code 0 as proof the work happened.
Reality
Prompting an LLM to self-audit works some fraction of the time, so one developer moved the check into a Stop hook that parses the session transcript and refuses the end of a turn when the audit keywords are missing.
Reality
Permission rules accumulate from "yes, don't ask again" answers into a repo-local file, and because matching runs deny then ask then allow with no specificity tiebreak, a policy that reads correctly can still stall an unattended run.
Reality
An adversarial self-audit check that blocks with exit 2 is a useful nag in a terminal. Under launchd it becomes a failed job. The fix here keys off session provenance rather than softening the rule.
Reality
A Windows box running about thirty unattended jobs reported success through three months of failures, because each exit code had to survive a VBScript launcher and a .cmd shim before the scheduler ever saw it.
Reality
The prompt only ever guarded sessions someone was watching, so the rules moved into a hook that runs in bypass and headless mode, refuses a force push by naming the sanctioned path, and self-tests on 59 synthetic calls.
Reality
A scheduled agent run reported four articles and zero replies for two days. The handle it was querying did not exist. The real numbers were five and two.
Reality