Safe Unattended Agent Automation with Five Laws

Description and use

The author describes a safe automation system using agents that run at night. Key principles include a runner managing locks, agents working in isolated branches without touching main, and generating a mandatory report even when no work is done. This prevents chaos and ensures transparency.

Problem

Unattended agents with write access pose a high risk of causing system instability or data corruption if they run without strict governance, leading to 'waking up to a mess'.

Outcome

The implementation of five strict laws allows for safe, unattended automation. The system successfully processes nightly queues without interfering with daytime work, ensuring that every morning produces a clear report of activities, skips, and required human interventions.

Workflow

• Schedule a nightly job to read a queue of pending work.
• For each target, the runner takes a claim (lock) on the target.
• The runner spawns a headless child agent in its own worktree.
• The runner refreshes the heartbeat on a timer while the child runs.
• The child performs work, commits to its own tree, or exits.
• The runner releases the claim always (on exit, crash, or kill).
• The runner pushes work to an isolated per-night branch (if enabled) and restores the previous branch.
• The runner generates an unconditional report detailing what happened, what was skipped, and what needs human attention.
• The developer reviews the report in the morning.

Evidence

Five nightly reports in, the streak is young but every morning has had one waiting, and that is the property I refuse to lose.

Source

Automation you have to check up on is just unfinished automation.

Source