Human In The Loop

Human-in-the-Loop

When to use

Use this skill when automation alone is insufficient because the task is ambiguous, high-impact, irreversible, privacy-sensitive, legally/materially significant, or requires domain judgement.

Objective

Introduce human judgement at the right decision points without slowing safe, reversible work unnecessarily. Approval must use architectural enforcement outside the model (tool dispatcher, policy engine or workflow gate), not prompt-only instructions the agent can ignore.

Procedure

1. Detect the need for human judgement using confidence, impact, reversibility and policy checks.
2. Pause execution before the risky action at a gate enforced outside the LLM (dispatcher or policy engine refuses execution until approval clears).
3. Present a concise decision packet: context, options, recommendation, trade-off, risk and required approval.
4. Resume only after clear approval or revised instructions recorded by the gate.
5. Record the decision and rationale in the implementation notes.

Intervention points

Ask for human input before:

• deleting, overwriting or migrating important data;
• changing security, authentication, permissions or secrets;
• making architectural choices with material trade-offs;
• accepting low-confidence model output;
• merging code when tests or validation are incomplete;
• exposing sensitive data to external services.

Decision packet template

Decision required: 
Context: 
Options: A / B / C
Recommendation: 
Main trade-off: 
Risk if wrong: 
Approval needed: 

Guardrails

• Do not bury approval requests in long prose.
• Do not ask for approval for trivial reversible steps.
• Do not proceed on implied approval for destructive or privileged operations.
• Do not rely on prompt text alone as the approval control; enforce the gate outside the model.
• Redact secrets and minimise personal data in review packets.

OWASP ASI09 controls

Use skills_docs/security/OWASP_ASI_CROSSWALK.md for the shared risk map. For ASI09 Human-Agent Trust Exploitation, approval requests must include evidence, uncertainty, consequences, alternatives and the exact action being authorised.

Related skills

• apply-laws-of-ai — escalation aligned with safety laws
• guardrails-safety-patterns — policy gates before human review
• agentic-ux-patterns — approval and evidence UI patterns

References

• OWASP Top 10 for Agentic Applications (ASI09 Human-Agent Trust Exploitation)
• OWASP AI Agent Security Cheat Sheet (HITL controls)

Verification

• [ ] HITL trigger is explicit.
• [ ] Options are clear and bounded.
• [ ] Recommendation and risk are stated.
• [ ] Approval action is unambiguous.
• [ ] Decision is recorded.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mmccalla
• Source: mmccalla/coding-agent-skill-library
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.