My agents run without permission prompts, so the brake moved into the hook

The author describes a shift in how they manage autonomous software agents, moving away from human-in-the-loop permission prompts toward a programmatic 'hook' system. Previously, permission prompts were ineffective for headless or background tasks, creating a security gap. The new approach uses a shell script hook that intercepts tool calls, evaluating them against a set of rules before execution. This system distinguishes between destructive actions that require redirection or refusal and routine tasks that pass automatically. By implementing a 'doctor' to validate the guard rules and ensuring the hook is universally applied across all sessions, the author has successfully automated safety checks. This transition allows agents to operate autonomously without stalling for human input, while maintaining strict control over sensitive operations like force pushes or recursive deletions, effectively replacing manual confirmation with a robust, rule-based gatekeeping mechanism.