An oversight posture where an AI agent operates autonomously while a human monitors from outside the execution path and retains the authority to intervene, halt, or roll back. Unlike Human-in-the-Loop (HITL) (where a human approves each action before it runs), on-the-loop keeps the agent moving at machine speed and only pulls a person in on exceptions, anomalies, or escalations. It's the posture most proven production agents settle into: fast, but with a real kill switch and monitored observability.
Пример
A SOC triage agent auto-closes low-severity alerts and enriches the rest around the clock. A human analyst watches a live dashboard, gets paged when the agent flags something it can't classify, and can freeze the agent's actions with one click. The analyst is on the loop — not approving each of the 4,000 daily alerts, but supervising the system that does.
Часто задаваемые вопросы
Is human-on-the-loop compliant with the EU AI Act?
Often yes. EU AI Act Article 14 requires *effective* oversight — the capability to understand, monitor, intervene, and halt the system — not necessarily pre-approval of every decision. For many high-risk agents, well-implemented on-the-loop oversight (real-time monitoring, a working kill switch, escalation, and audit logs) satisfies that, while the highest-risk cases still demand in-the-loop approval. Match the posture to the risk and level of autonomy, as the Article requires.
When should I move from in-the-loop to on-the-loop?
When you have evidence: an action class the agent handles correctly ~99%+ of the time with no severe misses, actions that are reversible or bounded in blast radius, and monitoring plus escalation actually wired up. Keep irreversible, high-blast-radius, and legally-required actions in-the-loop permanently.