# HITL / HOTL / HOOL: AI4SE's Spectrum of Human-Machine Oversight
Neither full HITL nor complete autonomy — true Harmony is an oversight spectrum routed by action risk: human-in-the-loop, on-the-loop, and out-of-the-loop each have their own place.
Three Modes of Oversight
| Mode | Definition | Typical scenarios |
|---|---|---|
| HITL | Agent proposes → blocks → human approves → executes | Merges, releases, deletions, compliance operations |
| HOTL | Agent executes autonomously, human oversees with override | Code exploration, test runs, draft generation |
| HOOL | Fully autonomous within policy + post-hoc audit | Auto-formatting, dependency updates, templated tasks |
Why Static HITL Isn’t Enough
Problems with full-time HITL:
- Humans become the bottleneck — waiting for approval at every step drains process efficiency
- Approval fatigue — too many trivial approvals means high-risk items get rubber-stamped too
- NIST AI RMF explicitly states: oversight should be “proportionate to risk”
Routing Strategy
if action.is_irreversible or action.is_compliance_critical:
→ HITL
elif action.is_exploratory or action.has_rollback:
→ HOTL
elif action.is_templated and action.has_gates:
→ HOOLReferences
- NIST AI Risk Management Framework
- Sheridan: Supervisory Control (1992)
- neurals.ca: HITL vs HOTL
- 2026 Taxonomy of Human-AI Interaction: HITL, HOTL, HIC