How do I implement human-in-the-loop for AI agents?

Human-in-the-loop (HITL) for AI agents means requiring human approval for high-risk actions while allowing low-risk actions to execute autonomously — and the key challenge is defining the boundary between "auto-approve" and "require human review" without creating bottlenecks.

HITL implementation patterns:

• Always-approve: Human reviews every action — safe but kills agent productivity. Only viable for <10 actions/hour
• Threshold-based: Auto-approve actions below a risk threshold (reads, small transactions), require approval above it (writes, large transactions, external communications)
• Anomaly-based: Auto-approve actions that match historical patterns, flag novel actions for human review
• Governance-based (Exogram's approach): Deterministic policy gates auto-approve compliant actions and block non-compliant ones — humans only review blocked actions that need exception approval

Exogram's approach is the most scalable: the 8 policy gates handle 99%+ of decisions automatically. Only truly ambiguous or policy-violating actions require human intervention. Blocked actions include the full context (proposed action, policy violation, state hash) so the human reviewer can make an informed decision in seconds. The agent continues processing other tasks while waiting for approval. No bottleneck.