Human-in-the-Loop

What is a Human-in-the-Loop

Human-in-the-loop is a workflow design where an agent's autonomy is deliberately interrupted at chosen points, most often before merging code, deploying, or taking some other action that would be costly to undo, so a person can review and explicitly approve before the agent continues.

The checkpoints are chosen, not automatic on every action. A well-designed human-in-the-loop workflow lets an agent research, draft, and iterate freely, then routes specifically the consequential decisions, approving a plan, reviewing a diff, confirming a deploy, through a person, rather than pausing constantly for things that do not actually need it.

This sits opposite AFK work on the same spectrum: AFK minimizes checkpoints in favor of speed and unattended progress, human-in-the-loop maximizes them at the specific points where getting it wrong is expensive. Most real workflows land somewhere between the two, autonomous through the exploratory and reversible parts, gated at the parts that are not.