Human-on-the-loop

An oversight pattern where the AI agent acts on its own while a person monitors its work and can intervene when something goes wrong.

Human-on-the-loop is an oversight pattern where the AI agent acts on its own while a person monitors its work and can intervene when something goes wrong. The person is a supervisor who can step in, not a gatekeeper who approves every step. An agent that processes routine expense reports might run on its own all day while a finance manager reviews a daily summary and flags anything unusual. This differs from human-in-the-loop, where every action waits for explicit approval.

Builder example

This pattern captures the speed benefit of automation for lower-risk tasks where requiring approval at every step would eliminate the time savings. An email-drafting agent that queues messages for batch review once a day is far more useful than one that interrupts you for every message. The key design decision: what triggers a human alert? Set thresholds on cost, sensitivity, confidence, or novelty so the supervisor sees what matters.

Common confusion: Monitoring and approval are different levels of oversight. If a mistake would be expensive or irreversible (sending money, deleting data, contacting a customer), human-in-the-loop approval is safer. Reserve on-the-loop for actions you can review after the fact and undo if needed.