Human oversight and agent autonomy: HITL, HOTL, HOOTL

They are three points on the spectrum of how much a human is involved in an agent's actions. Human-in-the-loop (HITL) means a person approves each critical action before it happens, so nothing consequential runs without an explicit human yes; it gives maximum control at the cost of throughput. Human-on-the-loop (HOTL) means a person supervises the agent as it works and can intervene or stop it, but does not sign off on every individual step; it suits medium-risk work where constant approval would be impractical but unsupervised action would be unwise. Human-out-of-the-loop (HOOTL) means the agent acts fully autonomously, with only after-the-fact review, which maximises throughput and risk together. The crucial point is that these are not a property of the whole system but of each action: a well-designed agent runs different actions at different levels, autonomously drafting a reply while requiring approval to send money. Treat the three as a dial you set per action, not a single setting you choose once for the agent.

Match autonomy to the risk and reversibility of each action, not to the agent as a whole. The two questions that decide it are how bad the worst outcome is and whether you can undo it: an action that is irreversible or high-impact — deleting data, moving money, sending external communications, anything you could not take back — should stay human-in-the-loop regardless of how capable the agent seems. Low-stakes, reversible, well-bounded actions are the candidates for higher autonomy, and even then you earn the move with evidence rather than assertion: evals showing reliability on the task and monitoring that would catch a regression. A useful test is to ask what happens on the agent's worst day, not its average one, and to set the level so that the worst day is survivable. Start conservative, raise autonomy action by action as the data supports it, and keep the irreversible actions gated for far longer than feels necessary, because that is exactly where an unattended mistake becomes an incident.

It is rolling out autonomy in stages so that real consequence is added only after the previous stage has proven safe, rather than flipping an agent to full autonomy on launch day. A typical progression runs dry-run, where the agent proposes actions but executes nothing; read-only, where it touches real systems but cannot change them; simulation, where actions run against a sandbox; staging, where it takes limited real actions under heavy monitoring; and finally production, with autonomy scoped to exactly what the evidence supports. The point of the sequence is that each stage earns the next: you only loosen control when the data from evals and observability shows the current stage is behaving, so you are raising autonomy on evidence rather than optimism. It also gives you a natural rollback — if a stage misbehaves, you drop back a level rather than ripping the agent out of production. Graduated deployment is the operational counterpart to the oversight dial: it is how you move along the spectrum deliberately instead of gambling on a single launch.

For high-risk AI systems, the EU AI Act requires meaningful human oversight — measures that let designated people understand the system, monitor its operation, and intervene or stop it — and the emphasis is on meaningful. The bar it sets is real oversight, not a token possibility: saying that a human could theoretically intervene does not satisfy it if there is no actual, designed checkpoint, no one whose job it is to watch, and no practical way to halt the system in time. In agent terms, that translates into the controls this discipline already describes: approval gates on consequential actions, a person genuinely able to supervise and stop the agent, and a kill switch that works. It also expects you to know and record which person or team is accountable for each system. The practical move is to map each agent's autonomy level to its risk tier, put a real human checkpoint where the risk warrants one, and keep the audit trail that shows oversight actually happened. Treat the legal requirement and the engineering best practice as the same thing, because here they converge.