Human review & escalation
Escalation is a product decision, not an accident. Use structured tasks when confidence is low, policy blocks automation, or the world requires accountable judgment.
When to route to humans
- Identity, payments, or safety outcomes that should not be self-certified by a model.
- Physical-world checks: on-site verification, deliveries, or in-person coordination.
- Low-confidence branches where the cost of error exceeds the cost of human latency.
- Anti-abuse and “prove you are human” checkpoints in agent-mediated flows.
Payload discipline
Include enough context for a human to decide without exposing secrets. Prefer references (ticket ids, signed links) over dumping full customer records into free text unless policy allows.
Closing the loop
After completion, reconcile your workflow: update CRMs, release holds, or resume agent plans. See Task lifecycle for the API side.