What is human-on-the-loop?

Human-on-the-loop is a design in which a system acts on its own while a person supervises it: watching what it does, sampling its work, receiving alerts, and able to intervene or stop it. The person does not approve each action beforehand. It suits routine, reversible work done at volume, where reviewing every item would cost more than it protects.

Why does human-on-the-loop matter once agents act?

Most of what a mature agent does should not wait for a person. Updating addresses, answering routine questions, applying exact-match payments: requiring approval for each would bury reviewers and teach them to click through. Human-on-the-loop lets the agent do that work while keeping a person responsible for how it goes.

The risk is that "on the loop" becomes "nowhere near the loop". Supervision that nobody actually performs is not oversight. A human-on-the-loop design needs tools that make supervision possible: a view of what the agent did, signals when something looks wrong, and a way to stop it quickly. The design principles are part of agent governance is a bound, not a model.

How does human-on-the-loop work in practice?

A working on-the-loop setup has five parts:

• Bounds that hold without the person: the agent can only act within its role, depth, sources and limits.
• A view of activity: the supervisor can see what the agent did, by capability and verdict.
• Sampling: a share of allowed actions is reviewed after the fact, with findings recorded.
• Alerts: unusual volumes, rising refusals or a run of corrections trigger a look.
• A stop: a named person can pause a capability or the whole agent at once.

The bounds are what make this safe. Because each action is checked at the moment of action, the supervisor does not need to catch every mistake; they need to catch patterns. Actions above the agent's depth or the customer's limits are held and go to a human gate instead.

What does that look like in an example?

A telecoms provider's agent handles SIM replacements for customers who have passed identity checks. It executes replacements at depth 3 without approval. A fraud team supervisor watches a dashboard of replacements by hour and region. One evening, the count from one region triples within an hour. The supervisor pauses the capability, which moves all new replacement requests to held, and the team finds a wave of fraudulent requests using stolen identity details. No individual replacement looked wrong; the pattern did.

How is human-on-the-loop different from human-in-the-loop?

In human-in-the-loop, a person approves each decision before it takes effect. In human-on-the-loop, the system acts and the person supervises. In-the-loop is stronger per decision and slower; on-the-loop scales and relies on good bounds and good monitoring. For agents, held actions are in-the-loop and allowed actions are on-the-loop, and most agents use both across their capabilities.

Neither design alone fits every situation, which is the point of should every AI agent have a human in the loop? The legal expectation, for example under the EU AI Act's human oversight requirements, is about oversight being effective rather than about one design, as discussed in human-in-the-loop vs human oversight under the EU AI Act.

Where Pryme fits

In Pryme, capabilities certified at depth 3 or 4 act within bounds without approval, and every verdict, allowed or not, is filed. That gives supervisors a complete record to sample and watch. Actions above the customer's limits are held for a named person, so one agent works on-the-loop and in-the-loop at once, by capability.