Human-on-the-loop
- Agents
- Standard term
Human-on-the-loop is an oversight setup where an AI agent acts without waiting for approval, while a person watches its work through alerts, summaries, or spot checks and can step in.
The person needs signals that reveal significant problems and a way to intervene while intervention can still help. Logs, alerts, and samples serve different purposes; their timing must match the consequences. Monitoring after an irreversible effect may support recovery without preventing it.
Where you’ll meet it
Your company's help-desk software adds an AI feature that answers routine customer questions on its own. You get a daily digest of its replies and an alert for anything it flags. A week in, the digest shows it offered a customer a refund the policy doesn't allow.
What to do
Pause automatic replies about refunds and contact the customer about the unauthorized offer. Trace whether the error came from the policy source, its interpretation, or the sending rules before changing the workflow. Keep refund offers under review until a checked control addresses that failure.
Easy to mix up
Human-in-the-loop work moves only as fast as a person gets to each request. Human-on-the-loop work moves at the agent's own speed, with a person watching and able to step in.
Check yourself
Which setup makes correcting mistakes after an agent acts more practical?
Go deeperHard limits and a short alert list keep a watching person effective
Hard limits complement both approval and monitoring. A spending cap or a ceiling on messages per hour can contain effects before a person has time to respond, provided it is enforced outside the model. Reaching the limit may stop the run or route it for a decision.
Useful alerts identify a concrete condition and an action someone can take, such as investigating an unusual charge or an unfamiliar request type. Frequent low-value alerts can bury important ones. Put routine information in a digest when it does not need an immediate response.
Writing on autonomous systems, including debates about autonomous weapons, uses “human on the loop” for supervision with a way to intervene. It also uses “human out of the loop” for operation without human supervision.
Sources
- LLM06:2025 Excessive Agency, genai.owasp.org
- Google SRE: Monitoring Distributed Systems, sre.google
- Losing Humanity, www.hrw.org