Guide · agent governance
Human in the loop vs on the loop: pick per action, not per system
The two phrases describe where the human sits relative to the action — a precondition, or a monitor. Choosing one posture for an entire agent is the mistake; consequence decides, action by action.
Three postures, one difference
In the loop. The agent stops and cannot proceed until a human decides. The human is a precondition. Nothing consequential happens while everyone is asleep.
On the loop. The agent acts; a human supervises and can intervene or reverse. The human is a monitor. Work continues overnight, and so do mistakes.
In command. The human sets and can revoke the mandate — which systems, which actions, what limits — without reviewing individual acts. This is governance, and it is what makes on-the-loop defensible: somebody owns the boundary even when nobody reads each line.
The question that decides it
Not “do we trust the agent?” but: if this action is wrong, what does it cost to undo?
A sandbox edit costs nothing — the container is thrown away. A pushed branch costs a revert. A sent email costs an apology and cannot be recalled. A refund is money that has moved. Reversibility, not confidence in the model, is what should set the posture.
The failure mode nobody admits
On-the-loop assumes a human is watching. In most deployments the dashboard exists, the alerts route to a channel, and nobody looks until something breaks. That is not supervision — it is autonomy with a monitoring bill.
The honest test is temporal: will a person notice inside the window where this is still reversible? If an agent can send two hundred emails at 3am and the first human glance is at 9am, the email action is not on-the-loop. It is unsupervised. Either gate it or accept what it can do unattended.
Why per-action beats per-system
Teams tend to pick one posture and apply it everywhere, then discover both extremes are unusable. Full in-the-loop means approving the agent’s every thought, so people stop using it. Full on-the-loop means the first genuinely bad action reaches production, so people stop trusting it.
Splitting by consequence keeps the agent autonomous where autonomy is free — reading, planning, editing in an isolated workspace, running tests — and puts a human exactly where the cost lands. In practice that is one decision per consequential act, which is a few seconds against a change that reaches customers.
FAQ
- What is the difference between human in the loop and human on the loop?
- In-the-loop means the action cannot happen without a person: the agent stops, a human decides, then it proceeds. On-the-loop means the agent acts on its own while a human supervises and can intervene or reverse. The distinction is whether the human is a precondition or a monitor — in-the-loop trades speed for control, on-the-loop trades control for speed.
- What does human-in-command mean?
- A third posture, from EU AI Act discussions: the human does not review individual actions but sets and can revoke the agent's mandate — what it may do, in which systems, within what limits. It is governance rather than supervision, and it is what makes on-the-loop safe, because someone owns the boundary even when nobody reviews each act.
- Which should I use for an AI coding agent?
- Both, split by consequence. Planning, editing in a sandbox and running tests should be on-the-loop — an agent that asks permission to think is useless. Anything that leaves the sandbox (pushing a branch, opening a pull request, touching production) should be in-the-loop. The mistake is choosing one posture for the whole system rather than per action.
- Is human-on-the-loop safe for AI agents?
- It is safe when two things are true: the action is reversible, and someone is actually watching. Most on-the-loop deployments fail the second test — the dashboard exists and nobody looks at it, which is really unsupervised autonomy with a monitoring bill attached. If nobody will notice within the window in which damage can be undone, treat the action as in-the-loop.
- Does human-in-the-loop slow agents down too much?
- Only when applied indiscriminately. Gate the actions that leave the sandbox and the agent still works autonomously for the ninety-plus percent of its runtime spent reading, planning, editing and testing. The cost of a gate is one decision per consequential act — measured in seconds, against a change that reaches customers.
see the whole flow, live
Everything on this page — the queue, the evidence, the hold-to-approve, the audit trail — is working in the VIVUU terminal as an interactive preview on realistic mock data. No sign-up.