HITL · HOTL · HOOTL
Three substantively different bets about where the human sits relative to an AI system producing work autonomously.
The acronyms have to do a lot of work in the next few years. Pricing, regulation, accountability, error handling, and market positioning all change as you move along the progression. The bets are not interchangeable; a product built for one is the wrong product for another.
HITL — Humans In The Loop
The human is a step in the iteration mechanism. The agent proposes; the human accepts, rejects, or edits; the loop advances one ratified action at a time. The human's attention is the throttle.
Cursor, Cline, Continue, Copilot, and Claude Code are HITL tools. Each is good at being good HITL. The safety story terminates at "did the human approve it?" Implicit assumption in most 2026 safety frameworks.
HOTL — Humans Outside The Loop
The human is the supervisor. The loop runs autonomously, internalizes its own rework through council reviews, audits, verdicts, and mutators, and produces artifacts the human reads to verify what happened. The human sets direction, ratifies at boundaries, and owns the final work product. The human is the composer; the runtime is the orchestra.
HOTL is a real and viable category for AI coding work and operational decision-making where the human cannot be the per-step ratifier without becoming the bottleneck. The composer framing is what makes HOTL safe: the human chooses what gets played, what gets shipped, and what gets killed.
HOOTL — Humans Out Of The Loop
The human is the signal. They stop stating intent; intent becomes a pattern the agents read from context — chat history, codebase shape, behavior over time, adaptive memory, market signal, ambient observation. The agents infer need, propose work, build, measure outcomes, and adjust.
HOOTL is the most autonomous of the three and the most legally and ethically exposed. When the human is not the iteration mechanism and not the per-step ratifier, the question of who is accountable when something goes wrong becomes substantively harder.
Most safety frameworks in 2026 implicitly assume HITL. Under HOOTL, that assumption fails. The eight HOOTL Safety Principles describe what a HOOTL system needs to be to be accountable.
The progression is also a research agenda
Each step requires capabilities the previous step did not need. HITL requires nothing of the agent's substrate; the human is the substrate. HOTL requires the agent to internalize rework — councils, verdicts, audits. HOOTL requires the substrate to infer intent reliably, to remain externally interruptible, to carry provenance through emergent autonomy. None of these are solved problems.
For the full taxonomy treatment including legal, ethical, and commercial implications, see the canonical text on GitHub.