Back to full roadmap
topiccore
Human-in-the-Loop Patterns
Approval gates, plan review, escalation — agent + human = safest production.
3 hours1 resources1 prereqs
Patterns:
- Plan review — agent generates plan, human approves/revises, then execute
- Action approval — approval before each destructive action (capability gating above)
- Escalation — low confidence or policy-violation suspicion → escalate to human
- Periodic checkpoint — long-running agent (hours/days) asks "shall I continue?" at intervals
- Post-hoc review — agent ran autonomously, results sampled + human review (10% sample)
LangGraph interrupt(): native HITL primitive. Agent resumes from saved state.