Human oversight: human-in-the-loop

Maps to: Human oversight: human-in-the-loop (interrupt/resume).

Scope

Pausing execution for approval, edit, rejection, or clarifying answers; resuming with structured payloads without holding workers for unbounded time.

Design questions

• Which tool calls require mandatory review versus conditional review by risk score?
• Payload shape surfaced to reviewers (diffs, previews, proposed arguments).
• Timeout, escalation, and default-deny behavior when humans do not respond.
• Parallel interrupts across branches and partial resume ordering.

Tradeoffs

• More gates increase safety and latency; fewer gates increase incident rate.
• Resume values that accept arbitrary JSON enable rich edits but complicate client UX.
• Dynamic interrupts in tool code travel with tools but scatter policy.

Evaluation hooks

• Approve, edit, and reject paths alter downstream tool execution correctly.
• Multi-day pause frees compute; resume continues from same logical step.
• Audit log ties human decision to run id and checkpoint.

Reference notes

See LangChain runtime article (HITL interrupt figure).