Human-in-the-Loop (HITL) — AI Glossary

Reference details

Related terms

• AI Agent
• Guardrails
• Agentic Loop

How it works

1. The agent's run must be suspendable at a defined point, which means its state has to be serialisable rather than held in a process's memory.
2. That state is written to durable storage with an identifier, and the run is resumed from it — possibly in a different process, on a different machine, hours later.
3. The approval request leaves through a separate channel entirely, since the reviewer is not watching the execution and the decision arrives asynchronously.
4. Resumption re-enters the loop with the decision appended as an observation, so an approval and a rejection are both just input the model reasons about rather than control flow.
5. A timeout policy is required because a pending approval that never arrives is a suspended run holding its state indefinitely, and the default of waiting forever is a resource leak.
6. The interrupt point also has to precede the action rather than wrap it: a tool that has already dispatched a request cannot be un-dispatched by a rejection arriving afterwards.

Example: Gate the irreversible step, not every step

An agent that can read a calendar, draft a reply and send an email has one action that cannot be undone. Gating all three means a human confirms three times per task; across twenty tasks a day that is sixty approvals, and the reviewer starts approving without reading — the automation now has the appearance of oversight and none of the substance. Gating only the send is one approval per task, on the step that matters, and it stays meaningful. Choose interrupt points by reversibility and blast radius, not by uniformity.

Commonly confused with

Human-in-the-loop is not human-on-the-loop. In the loop, a human approves before the action proceeds. On the loop, the system acts and a human monitors and can intervene after. The second scales far better and is the weaker guarantee; saying which one you mean is the whole design decision.

When to use it

Reach for it when:

• Irreversible or externally visible actions: payments, sends, publishes, deletions
• Low-confidence cases routed for review, with the threshold tuned on measured error rates
• Early deployment, as a source of labelled corrections that later become evals

Reach for something else when:

• High-volume reversible steps, where approval fatigue destroys the review's value
• As a compliance fig leaf when the reviewer lacks the context or time to genuinely refuse
• Latency-critical paths, unless the interrupt is designed to be asynchronous

Referenced by

This term is used by the following ModelRefs references:

• file write tool
• tool approval gateway

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Human-in-the-Loop (HITL) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path