How to Build a Good Human-in-the-Loop for AI Coding Agents

Short summary

Effective human-in-the-loop for AI coding agents requires grading each action by reversibility, blast radius, and stakes (G0-G3), then matching controls to the grade. Research shows humans only catch 9-26% of harmful mid-task actions, making per-action approval prompts theater for irreversible operations. The framework lets trivial actions run, makes low-risk edits undoable, requires review for merges, and outright prevents destructive shell commands that undo features can't cover.

• •Grade agent actions G0-G3 by reversibility × blast radius × stakes, then match controls accordingly
• •Humans catch only 9-26% of harmful mid-task actions — approval prompts are ineffective for irreversible operations
• •Shell side-effects (rm, mv, cp) sit outside undo boundaries and must be prevented, not gated behind prompts

Generated with AI, which can make mistakes.