Pattern · Agents
Human approval for actions
Pause before costly, irreversible or external actions; show a person the action and evidence; resume with their decision.
Say it in the interview
I would let the agent act alone on reads and small reversible steps, and pause on anything costly or external: save state, show a person the action with its evidence, and resume with their decision. I would log the decision time to tune the thresholds.
When to use it
- Refunds, payments, emails to customers, record changes, deletions
- Regulated decisions
How it works
- Agent proposes an action → Risk rules
- Risk rules → read, or small and reversible → Run it
- Risk rules → costly or external → Save a checkpoint and pause
- Save a checkpoint and pause → Approval card (action, reason, evidence)
- Approval card (action, reason, evidence) → approve, edit or reject → Resume from the checkpoint
Trade-offs
- Safety and accountability
- Slower; needs people on call and reminders
Pitfalls interviewers probe
- Approval on everything, so people click yes without reading
- No timeout or escalation when nobody answers