Design where a person approves, edits, or rejects high-risk agent actions before they commit—common for deploys, deletes, payments, or production config changes.
HITL is not a compromise on autonomy; it is guardrail infrastructure. Agents propose; humans confirm when blast radius is high or policies require it.
In the sample curriculum
Tooling sections describe approval gates, dangerous tiers, and middleware that pauses execution until a reviewer responds—those are concrete HITL implementations.