Human-in-the-loop

Human-in-the-loop means a person stays part of the run: the agent pauses for a permission request, asks a question when the task is ambiguous, or waits for an answer before it commits to a direction. The human is a step in the loop, not just a reader of the output at the end. In most coding harnesses this is the default: the agent proposes an edit or a command, you approve or redirect, it continues.

The trade-off is throughput for control. Every pause costs your attention, and if you rubber-stamp fifty prompts in a row you get the cost of supervision without the benefit. The opposite failure is just as common: leaving the loop for an hour and coming back to polished work that took a turn you'd have stopped had you been watching.

Use it deliberately. Stay in the loop when the task touches money, data, other people's systems, or anything you can't roll back; keep the interruptions meaningful by pre-approving the safe, boring tool calls in your permission mode. Move toward AFK runs only once automated checks can catch what you would have caught by watching.