Human in the Loop

Human-in-the-loop is a key safeguard at critical junctures: some judgements require human values and expertise.

Two triggering situations:

  • failure threshold exceeded — escalation to a human;
  • high-risk operations — confirmation before execution.

Timeout strategies and default behavior are needed (“no response in 5 minutes — conservative strategy”).

Human approvals and rejections with justifications form feedback data for continuous evolution.

Related: Fences, Agent Continuous Evolution, Proposer-Reviewer