Building Agents
Human-in-the-Loop Patterns for AI Agents
Where approval, review, correction, and escalation create the most value in agent workflows.
Approval before consequence
Pause before actions that are expensive, public, destructive, legally meaningful, or difficult to reverse. Show the proposed action, supporting evidence, and expected effect rather than a generic confirmation dialog.
Review after reversible work
For drafts, classifications, or queued changes, asynchronous review can preserve speed. Capture corrections as evaluation data instead of silently discarding them.
Escalation is a success path
Agents should recognize missing authority, conflicting evidence, and cases outside their competence. A clean handoff with context is better than a confident but unsupported attempt.