Here's a quiet shift happening in almost every organization running AI agents, usually without anyone deciding it. Human-in-the-loop, where a person approves each consequential action before it happens, is becoming human-on-the-loop, where the person watches the system run and steps in when something looks wrong. Mostly nobody announces this. The agent takes on more actions, per-action approval becomes the bottleneck, approvals get batched, then sampled, then skipped, and one day you look up and the human isn't in the loop anymore. They're near it.
Let me say the uncomfortable part first: the shift is necessary. I argued back in Issue 002 that "a human reviews everything" is advice nobody can run at scale, and agent scale makes it arithmetic. An agent taking a thousand actions an hour cannot wait on a person per action. Even the people writing the standards concede the point: the NIST agent standards work names the practical impossibility of meaningful real-time human oversight of long-running autonomous processes as one of the core properties that make agents hard to govern. So the question isn't whether you move on-the-loop. It's whether you notice what you traded away when you did.
What human-on-the-loop actually gives up
Because it does trade things away, and the trades have names. On-the-loop is not in-the-loop with less work. It's a different control model.
You give up the veto before the act. In-the-loop, nothing happens until someone says yes: the mistake is prevented. On-the-loop, the mistake runs first and gets caught after. Monitoring doesn't stop an error. It shortens how long the error lives, which for a reversible action is fine and for an irreversible one is no control at all.
You give up forced attention. Approval made a person look at every action, whether they felt like it or not. A dashboard makes nobody look. Monitoring attention decays, alert floods teach people to ignore alerts, and the system's own reliability works against you: the better it runs, the less anyone watches.
You give up the intervention window, or you pretend you have one. "We can step in" assumes the action waits long enough for a person to notice, judge, and act. If the agent executes in seconds, an intervention story measured in minutes is fiction wearing a control's name.
You give up practice. Reviewers who no longer make the calls slowly lose the ability to judge them. The day the escalation finally lands on a human is the day the judgment has quietly atrophied from disuse.
And you give up a clean answer for the miss. "A human approved it" was the accountability anchor. When everyone watches and nobody approves, an error belongs to no one in particular, which in practice means it belongs to whoever is worst at politics.
None of this is an argument against on-the-loop. It's an argument against drifting into it. Every one of those trades can be compensated, deliberately: keep pre-approval for the irreversible actions and let only the undoable run ahead of you. Alert on deviation rather than volume, and test whether a planted anomaly actually gets noticed. Time your intervention window honestly, and gate before execution anything you couldn't stop in time. Route a share of real decisions to humans on purpose, purely to keep the judgment muscle alive. And put a name on the watching, per system, so a miss lands somewhere.
Notice what those compensations have in common: every one is a decision, made in advance, about a specific class of action. That's the actual discipline. In, on, or out of the loop was never a system-wide setting. It's a choice per action class, set by what a mistake costs and how hard it is to undo. The stakes pick the loop. The drift I opened with, approvals batched, sampled, skipped, is what it looks like when workload picks the loop instead.
So run the honest audit. List what your agents can do, mark what's irreversible, and ask where a human really sits for each: in the loop, on it, or just nearby in the org chart. The first two are control models. The third is a story, and it holds up exactly until the first miss goes looking for an owner.