Some actions are too consequential, too irreversible, or too novel to let an agent take alone - even a well-tested one. HITL checkpoints pause the run at a chosen point, write a HITLReview row, and wait for a human to approve, reject, or amend before the agent continues.
When to use HITL
- Approval-amount thresholds - invoice over $10k, refund over $500
- Irreversible actions - posting to ERP, sending external email, deleting records
- Regulatory requirements - GDPR data deletion, financial reporting submissions
- High-novelty inputs - the agent's confidence score is below threshold
- Anything legal or compliance has asked you to gate
Adding a checkpoint
Agent Builder > Guardrails tab > Require human review. Pick the trigger from the dropdown - Always (every run pauses), Condition (e.g., invoice amount over your approval limit), or Confidence (the agent's own confidence below a threshold). Optionally name a reviewer role or specific user and set the expiry window.
What a pause looks like
The runtime creates a HITLReview row with status=pending_review and an expires_at timestamp. The run pauses at the node. The configured reviewer (or anyone with the named role) gets an in-app notification + email with a one-click link to the review screen.
Responding
- Reviewer opens the HITL Reviews queue (sidebar > HITL Reviews)
- They see the agent's reasoning so far, the inputs, the proposed next action, and an Approve / Reject / Amend control
- Approve resumes the run from the checkpoint with the agent's proposed action
- Reject ends the run and writes the reason as the final output
- Amend lets the reviewer edit the agent's proposed action before resuming
Escalation
Configure a level-2 reviewer rule that fires if the level-1 reviewer misses the deadline. Common pattern: invoice approval auto-escalates from Finance Manager to CFO after 24 hours. The escalation chain is configurable to as many levels as you need.