Human-on-the-Loop vs Human-in-the-Loop: Picking AI Oversight

Human-in-the-loop vs human-on-the-loop vs out-of-the-loop: what’s the difference?

Model	When the human acts	Speed	Fits
Human-in-the-loop	Before each risky action	Slow, limited by reviewers	Payments, external messages, irreversible steps
Human-on-the-loop	During and after, on alerts or samples	Fast	Frequent, reversible, low-impact work
Human-out-of-the-loop	Not in real time	Fastest	Narrow, well-tested tasks

Most real systems mix the three by action, not by system. One agent can draft replies on the loop and send refunds in the loop, which is the idea behind agent autonomy levels.

What does human-on-the-loop require to be real?

The supervisor has to see what happened, in time, and be able to stop it. That means logs a person can read, alerts on unusual behavior through AI observability, sampling of routine output, and a working kill switch. A person who cannot see or stop the system in time is on the loop in name only. EU AI Act Article 14 asks the same of high-risk systems: people overseeing them must be able to monitor operation, interpret output, and interrupt it through a stop button or similar procedure (Regulation (EU) 2024/1689, read 12 November 2026).

How do we run our own agents on the loop?

Our scheduled agents, such as the one that drafts entries for this glossary, run unattended and nobody approves their steps. They are on the loop by design: each run works on its own branch, opens a pull request, and the change merges only after automated tests, builds and an independent review pass; the run then posts a summary to a team channel that a person reads. Which actions an agent may take alone follows three written tiers: decide and record it, propose and stop at something reviewable, or ask for an explicit yes. Anything in the ask tier is out of scope for an unattended agent by construction, so it prepares the artifact and leaves the decision on a board for a person. The loop is defined by what the agent may not do alone.

Written with AI assistance and reviewed by Simon Budziak. The production notes come from systems Soba Labs builds and runs.

Frequently asked questions

What is the difference between human-in-the-loop and human-on-the-loop?

In human-in-the-loop the system waits for a person to approve an action before it happens. In human-on-the-loop the system acts first and a person supervises, reviewing results and stepping in when something looks wrong.

When is human-on-the-loop appropriate?

When actions are frequent, low in impact and reversible, and when the supervisor can see what the system did and stop it in time. Payments, external messages and anything irreversible usually still need approval before they run.

What is human-out-of-the-loop?

It is full autonomy: the system acts with no person monitoring in real time. It suits only narrow, well-tested tasks where a wrong action is cheap and is caught by other controls.