Human in the loop

Human in the loop describes an AI workflow where a person reviews or approves before an action takes effect. In practice it is applied proportionately: low-stakes work runs unattended, medium-stakes work receives lightweight approval, and irreversible or externally visible actions require structured review.

Where the checkpoint sits matters more than whether one exists. Approving an agent's plan up front and then letting it run unattended provides less assurance than it appears to, because the plan is not where things usually go wrong.

The useful criteria for deciding are reversibility, blast radius, external visibility, and detectability. An action nobody would notice was wrong until an audit deserves review that an immediately self-correcting one does not.