Human in the Loop Escalation

Define the decisions an agent must hand back to a person instead of making alone

Use when: an agent can act on systems where a wrong call is expensive or irreversible

Avoid when: the workflow is entirely read-only and reversible

Procedure

1. List decisions with irreversible or external effect
2. Write the trigger for each: monetary, destructive, legal or scope crossing
3. Make the agent stop and state the decision it needs
4. Include the evidence and the options with their consequences
5. Record the human decision so it is auditable later
6. Re-check the list when the toolchain gains new capability

Signals

• irreversible actions have a named human approver
• the escalation message contains evidence, not just a question
• the decision is recorded with who made it
• new tools trigger a review of the escalation list

Patterns

1. Trigger list

delete, publish, or any target outside scope

2. Escalation message

decision needed: publish pack to production

3. Decision record

approved by operator at 2026-10-29T08:12Z

Pitfalls

• Escalating everything, which trains the human to approve without reading
• Escalating after acting rather than before
• Leaving the escalation list unchanged after adding a destructive tool

Done when

• every irreversible action has a named human approver
• escalations include evidence and are recorded

References

• NIST - human factors
• OWASP - secure design