How to prevent AI coding agents from making destructive changes without human approval

AI coding agents can edit files, run commands, and alter configurations in seconds, raising questions about authority rather than capability. A proposed governance framework classifies agent actions into low-impact reversible tasks, high-risk destructive operations, and undefined-authority situations that require a human approval gate. Before executing any hard-to-reverse action, the agent must document the impact, a rollback path, the justification, and explicit human sign-off — otherwise execution halts. The framework also enforces evidence-based success claims, meaning an agent can only mark a change complete when observable proof exists, not merely the absence of visible errors. A free Portuguese-language starter kit with four policy and checklist files has been published on GitHub, while a more comprehensive paid governance kit is available for R$39.90.