Agents / Standard term
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 in which an agent can act without waiting for approval while a person monitors summaries, alerts, or sampled runs and can intervene. A finance team might let an agent categorize routine expense reports, alert on unusual amounts or policy conflicts, and preserve the records needed to correct a classification. The person supervises the operating system rather than approving every step.
Builder example
The pattern is useful when actions can be observed soon enough for intervention to matter. Good monitoring needs a defined signal, an intervention mechanism, and a response time that fits the potential harm. Cost, data sensitivity, novelty, repeated failure, and disagreement among checks can trigger an alert. Automated containment should still stop effects that can spread faster than a person can respond.
Common confusion: A daily report alone does not create meaningful oversight if damage becomes irreversible within seconds. Human-on-the-loop can govern consequential work when limits, rollback, and alerts keep intervention effective; a blocking approval is appropriate when only a named person can authorize the action or no later correction can address the consequence.