Human-in-the-Loop (HITL) Patterns

Overview

The "Human-in-the-Loop" (HITL) skill provides a standardized framework for integrating human oversight into agentic workflows. By tiering actions based on risk, we ensure autonomy is utilized safely for low-impact tasks while maintaining rigorous human control over high-impact operations. This skill aligns with the C4 Runtime Protocol to prevent "black box" agent behavior.

When To Use

• Before performing destructive file operations (e.g., rm -rf, bulk overwrites).
• Before triggering deployment pipelines or infrastructure changes.
• Before making financial or security-impacting decisions.
• When an agent's confidence in a plan or outcome is below a critical threshold.
• When the user requires explicit auditability of agent intent.

Risk-Tiered Action Matrix

To manage friction, apply the appropriate gate based on the action category:

Tier	Category	Mechanism	User UX
0	Informational	Auto-Approve	Logged in background/console.
1	Standard Edit	Warn/Confirm	Inline diff preview + "Proceed".
2	Destructive/Critical	Hard Gate	Explicit approval required; blocked until sign-off.
3	Financial/Deployment	Multi-Factor/Formal Audit	Requires explicit sign-off via UI/Artifact link.

Checkpoint Design Patterns

1. The Plan Preview: Agents must expose the plan artifact before execution.
2. Structured Clarification: Use tool-based, multi-choice queries to reduce ambiguity, avoiding open-ended questions that fatigue users.
3. Delta Validation: Always present a diff or specific "what is changing" summary before writing.
4. Resumption Points: Save state tokens to ensure that if a session drops, the human can review the exact state, modify if needed, and resume.

Implementation Workflow

1. Identify Risk: Evaluate the proposed task against the Risk-Tiered Action Matrix.
2. Surface Intent: Generate an "Action Summary" using the ask_question tool or an artifact.
3. Await Authorization: Use non-blocking pauses (or asynchronous messaging) to wait for approval.
4. Execute with Trace: Perform the action under a logging wrapper that records the authorized state ID.
5. Verify: Immediately follow with a verification check to confirm the expected outcome.

C4 Integration

HITL checkpoints are first-class citizens in C4:

• Brainstorming Gate: Agents must have their proposed path confirmed by the user before transitioning to ready.
• Verification Gate: After execution, the agent must present the "result vs. expectation" for user sign-off to close the cycle.

UX Patterns

• Transparency: Clearly display why a decision was made.
• Micro-interactions: Keep confirmation dialogs context-aware (e.g., provide links to the specific files being modified).
• Rollback Safety: Always provide an "Undo/Rollback" pathway for any high-risk action taken.

Anti-patterns

• Over-gating: Asking for permission on routine/low-impact tasks (causes "alert fatigue").
• Under-gating: Proceeding with high-risk/destructive actions without explicit, documented user consent (unsafe autonomy).
• Vague Prompts: "Is this okay?" (inadequate description of change). Always use: "I am about to replace X with Y. Do you approve?"

References

• KILO-KIT C4 Runtime Protocol
• Systematic Debugging
• Verification Before Completion