HITL UI Patterns
Human-in-the-loop (HITL) UI patterns balance agent autonomy with human oversight, evolving beyond simple binary approvals into sophisticated collaboration models.
1. Core Interaction Patterns
- Approval Gate (Interrupt & Resume): The agent pauses and waits for user approval before executing a specific action.
- Confidence-Based Routing: The agent only requests help when its internal confidence score falls below a threshold.
- Collaborative Drafting: The agent presents a draft that the human can refine in an inline editor before the agent continues.
- Strategic Guidance (Fork in the Road): The agent asks the user to "Set Direction" at critical decision points rather than just approving a task.
2. UI/UX Design Principles
- Decision-Ready Narratives: Summarize the Intent, Action, and Impact rather than showing raw JSON.
- Irreversibility Highlighting: Use visual cues (e.g., red borders) for high-risk, non-undoable actions like data deletion.
- Active Engagement: Require the user to answer a question or highlight text before the "Approve" button activates to prevent rubber-stamping.
3. Risk Classification Tiers
- Low Risk (Reversible): Auto-execute + Audit Log.
- Medium Risk (Uncertain): Confidence-based review.
- High Risk (Irreversible): Mandatory Approval Gate.