Human-in-the-Loop (HITL)

Human-in-the-loop is a control pattern where the agent pauses at predefined checkpoints and waits for a human to approve, reject, or edit its proposed next action. The human is part of the loop, hence the name. Common HITL checkpoints: before sending external communication (email, calendar invite), before mutating customer data, before spending money, after detecting low confidence, when the agent enters a sensitive policy area. HITL is the standard pattern for high-stakes domains like legal, medical, and financial services where unattended autonomy is unacceptable.

The cost of HITL is latency. Every checkpoint adds human response time, which can be minutes or hours. Two design patterns reduce that cost without removing the safety. First, asynchronous approval: the agent surfaces the pending decision in a queue (Slack, email, dashboard) and continues with low-risk parallel work while waiting. Second, tiered checkpoints: only the riskiest decisions trigger a human, lower-risk ones use automated guardrails. LangGraph and Temporal both have first-class support for the pause-and-resume pattern HITL requires.

Use HITL whenever an action is irreversible, expensive, or compliance-sensitive. The cost of one human review is always lower than the cost of one bad autonomous action in those domains.

• LangGraph: Human-in-the-loop
• Anthropic: Building Effective Agents

Related Terms