Human-in-the-loop

Human-in-the-loop

Also called: 人在回路 · HITL · 人工审批

A design pattern that inserts human approval at key points in an agent's process — the mainstream safety strategy for enterprise deployments.

AgentproposesHumanapprovesActionexecutedgate on risky steps
Schematic (simplified)

Where the gates go

Not every step needs review — gates go before actions that are hard to reverse or outward-facing: payments, transfers, deleting data, sending email, publishing content, changing production config. The agent presents "what it intends to do" with the arguments, and only executes after a human confirms.

Related variants

  • Human-on-the-loop: the human does not approve each step but monitors in real time and can stop it.
  • Human-in-command: the human keeps authority over goals and shutdown, delegating routine execution.

Which to use depends on the risk and reversibility of a single action — in-the-loop for high risk, on-the-loop for low-risk high-volume work.

Common misconceptions

  • Human-in-the-loop is not "a weaker agent" — it is an engineering response to irreversible risk.
  • Requiring confirmation on every step destroys the agent's value; the skill is choosing the right gates.

Related terms

Sources

  1. Anthropic — Building effective agents: guidance on human oversight (2024)
  2. NIST AI Risk Management Framework (AI RMF 1.0)

Compiled 2026-08-29 · This glossary is compiled from public papers, official specifications, and common industry definitions, and is updated as the field evolves. Corrections welcome.