Humans are not just emergency brakes
Human-in-the-loop is often framed as safety approval, but it is also a way to inject judgment, taste, and missing context. Good agents know when a human answer is cheaper than another five tool calls.
The checkpoint should be specific. Ask one decision at a time, present consequences, and carry the answer forward in state.
Checkpoint design
Ask before irreversible actions. Ask when requirements conflict. Ask when the user likely has private context. Ask when the next step spends money or changes external state.
Do not ask vague questions like “what should I do?” Ask “Should I delete these 3 generated files, or keep them for review?”