ModelRefs / Human-in-the-Loop — Agent Pattern
Human-in-the-Loop — Agent Pattern
Pause the agent at destructive steps for human approve / reject / edit before continuing. The agent resumes from a checkpoint after the human responds.
Overview
When the agent proposes a destructive action (send email, charge card, delete file), it pauses and surfaces the action to a human via Slack, email, or UI. The agent resumes from a checkpoint after the human responds.
When to use it: Your agent takes destructive actions and you cannot tolerate silent mistakes.
Pattern details
| Pattern class | safety |
|---|---|
| Difficulty | intermediate |
| Autonomy | guided |
| Also known as | hitl, approval gate |
| Last reviewed | 2026-06-07 |
Known failure modes
- Stale checkpoint — World changed between propose and approve. Mitigation: Re-validate preconditions on resume.
- Approval fatigue — Humans rubber-stamp without reading. Mitigation: Show diffs + risk score; require explicit edits for high-risk.
When not to use it
- Approving in chat without persisting state.
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 — Agent Pattern.
Frequently asked questions
When should I use the Human-in-the-Loop agent pattern?
Your agent takes destructive actions and you cannot tolerate silent mistakes.
What are common failure modes of Human-in-the-Loop?
Stale checkpoint • Approval fatigue
Is Human-in-the-Loop production-ready?
Yes when paired with the safety controls and observability hooks documented on the pattern page.