Human-in-the-Loop Control
Human-in-the-Loop Control
Key jargon
| Term | Plain-language meaning |
|---|---|
| Human in the loop (HITL) | A workflow that intentionally requires human input at a defined point. |
| Interrupt | A durable pause that preserves state while awaiting a decision. |
| Escalation | Routing a case to a person or higher authority when automated handling is unsuitable. |
| Four-eyes control | Requiring a second person to approve a sensitive action. |
Key concepts
- Place human gates at consequential decisions, not at every harmless model call.
- An approval must show the proposed action, evidence, scope, and likely effect—not just an abstract yes/no prompt.
Concept map
flowchart LR
A["Agent proposes consequential action"] --> B["Persist interrupt and evidence"]
B --> C["Human approves edits or rejects"]
C --> D["Resume or close"]A human gate works only when the reviewer sees the proposed action, target, evidence, alternatives, uncertainty, cost, reversibility, and policy decision.
Gate types
| Gate | Purpose |
|---|---|
| Intake | Confirm task, authority, and data scope |
| Plan | Approve decomposition and intended tools |
| Exception | Decide on conflict, uncertainty, or scope expansion |
| Effect | Authorize a specific external change |
| Release | Accept quality/safety evidence before deployment |
| Incident | Stop, contain, rollback, and preserve evidence |
Approval fatigue
Many low-information prompts train people to click “approve.” Batch harmless actions, reserve gates for meaningful transitions, and make rejection safe.
Exercise
Design an approval card for sending an email: sender identity, recipient, subject, body preview, attachments, data classification, source of content, and undo/recall limitations.