HITL (Human-in-the-Loop Approval)

What Problem It Solves

For high-risk actions, “best-effort” automation is not enough. HITL adds a human approval gate:

• approve/deny tool calls (or full plans)
• collect clarifications (missing info, ambiguous intent)
• create an auditable decision trail

When to Use

• The agent can trigger irreversible actions (payments, deletions, email sending).
• You need operational control and accountability.
• You want a safe path to gradually increase autonomy.

Core Flow

flowchart TD
  A["Agent proposes action"] --> H["Request approval (HITL)"]
  H -->|approved| T["Execute tool / step"]
  H -->|denied| R["Revise / ask for clarification"]
  T --> O["Observation"]
  O --> A

Evolution Path

• Built on: Policy + Guardrails
• Next steps:
• Multi-agent handoff (triage to the right human role/team)
• Eval harness (ensure approval thresholds and risk logic stay stable)

Repo Reference

• Code: src/agent_patterns_lab/runtime/hitl.py
• Example: examples/66_governance_hitl_policy_guardrails.py
• Tests: tests/test_hitl.py