Human-in-the-Loop
Pause agent execution and wait for human approval before any dangerous action proceeds.
How it works
1
Agent sends event via SDK
The event is evaluated by the Rules Engine.
2
Rule matches → require_approval fires
The platform creates an approval request.
3
Notification sent
Email + Slack notification to the assigned reviewer.
4
Admin reviews in Dashboard → Approvals
Full context: event, metadata, risk score, agent info.
5
Admin approves or rejects
Decision recorded with timestamp and reason.
6
Agent receives response
SDK polling resolves with the decision.
Timeout Behavior
| Option | Behavior | Recommended for |
|---|---|---|
auto_block (default) | Action blocked automatically | Financial, delete, export |
auto_approve | Action proceeds with warning | Low-risk actions only |
escalate | Forwarded to next role | Enterprise critical |
Default timeout: 30 minutes. Reminder sent 5 minutes before expiry.
Escalation Chains
Business/Enterprise plans support multi-level escalation:
- L1: developer@company.com (15 min)
- L2: techlead@company.com (15 min)
- L3: cto@company.com (30 min)
- AUTO-BLOCK if no response