Human-in-the-Loop
An AI system design pattern in which human judgment is incorporated at designated checkpoints to approve, correct, or redirect agent actions before they become irreversible or cross a policy boundary.
Human-in-the-loop (HITL) is not a single mechanism but a spectrum of interventions. At the lightest end, a human reviews agent outputs after the fact (asynchronous review). In the middle, a human approves each plan before execution begins but lets execution run autonomously. At the heaviest end, every tool call requires explicit human confirmation—appropriate only for high-stakes, low-volume workflows.
Designing effective HITL requires identifying the right intervention points: not so frequent that humans become a bottleneck, not so sparse that errors propagate too far before detection. Common trigger conditions include: spending above a threshold, accessing sensitive resources, taking irreversible actions, or reaching low-confidence states.
In Paperclip's G0–G4 approval model, G4 is the human approval gate. Three channels surface the approval request (email magic-link, Slack/Teams button, Paperclip UI queue) to maximize the chance of a timely human response without requiring the human to monitor any single interface.