Supervise an autonomous coding agent by controlling its task, permissions, checkpoints, and definition of done. Autonomy should describe how long it can execute between reviews, not freedom from accountability. Begin with a bounded goal and an isolated, reversible workspace.
Specify the repository revision, allowed files and commands, network and credential access, time and cost limits, acceptance checks, and conditions that require a stop or human approval. Review the plan before high-risk work. Require checkpoints after discovery, before migrations or dependency changes, and after failed tests. Preserve logs, tool calls, diffs, and test results. Configure distinct states for waiting, failed, cancelled, and verified completion so silence or process exit cannot look like success.
Match supervision to risk. Documentation and focused tests may need only final review. Authentication, payments, production data, infrastructure, and destructive actions need tighter approval gates. Watch for scope growth, repeated retries, broad file churn, disabled checks, and assumptions presented as facts.
Start with short tasks and compare the agent’s reports with actual repository state. Increase duration or permissions only after the controls prove reliable. Good supervision makes intervention predictable: you know when the agent must ask, what evidence it must return, and how to stop or roll back the work.
