Glossary

Human-in-the-loop

The system cannot complete the loop without a person, by design.

Definition

Human-in-the-loop (HITL) describes a system design in which a person's judgement is a required step, not an optional one: the process pauses for human input, approval or correction before it can continue or take effect. The term originated in control and simulation engineering, spread to machine learning through data labelling and active learning, and now describes approval gates in AI agent systems.

Abbreviation: HITL
Origin field: Control and simulation engineering
Sibling terms: Human-on-the-loop, human-in-command

Where the term came from

The phrase comes from control engineering and military simulation, where a human-in-the-loop simulation is one that cannot run without a live operator making decisions inside it. Machine learning borrowed it for training pipelines where people label data, correct model outputs, or choose which examples the model should see next.

AI agents inherited the term with a narrower meaning: a checkpoint where the agent must stop and get a person's decision before proceeding or before an action takes effect. Regulatory frameworks, including EU AI Act provisions on human oversight of high-risk systems, have since given it legal weight in some contexts.

Commonly confused with

These three are genuinely different levels of oversight, and the distinction matters in policy documents.

Term	The human's position	If the human does nothing
Human-in-the-loop	A required step inside the process	Nothing happens. The process is blocked.
Human-on-the-loop	Monitoring, able to intervene or veto	The system proceeds on its own.
Human-in-command	Setting policy and scope, not individual decisions	The system operates within the boundaries already set.
Fully autonomous	Outside the operating loop entirely	The system acts and reports afterwards, if at all.

Where the checkpoint usually goes in agent systems

Before a consequential action
Sending an external email, moving money, deleting records, publishing. The cost of the action is irreversible, so the approval sits in front of it.
At delivery
The agent produces work and a person accepts, rejects or asks for changes. The most common gate, and the cheapest to operate.
On low confidence
The agent escalates only when it is uncertain or when the case falls outside its brief. Efficient, and dependent on confidence estimates being honest.
On a sample
A percentage of completed work is reviewed to detect drift. Used where per-item review would be more expensive than the work itself.

The failure mode nobody plans for

A review gate that always approves is not oversight, it is a click. Rubber-stamping is well documented in automation research under automation bias: people asked to check machine output repeatedly, and rarely finding a fault, stop checking properly. A gate placed on every trivial output produces exactly that outcome and then fails on the one case that mattered.

The practical answer is to place fewer gates on higher-stakes decisions and give the reviewer something to check against: acceptance criteria written before the work started, and a record of what the agent actually did.

Acceptance criteria

Written before the work, checkable after it, and binary either way.

Autonomous agent

Autonomy is a range, and the interesting question is where the boundary sits.

Work log

The record that makes an unwatched run reviewable afterwards.

Delivery comment

Work arrives where the task already lives, attributed and reviewable, and the task stays open.

Agents deliver, humans close

One rule holds the whole product together, and it is a rule about who is allowed to say finished.

Agentic project management

The line is whether the agent holds the work item or only summarises it.

Questions people ask

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

In-the-loop means the process cannot continue without the person: if nobody acts, nothing happens. On-the-loop means the person supervises and can intervene, but the system proceeds by default. The distinction decides what happens when a reviewer is asleep, on holiday, or simply not looking.

+Does human-in-the-loop slow everything down?

It adds latency at the gate, which is the point. The design question is not whether to have gates but where to put them: on irreversible actions and final deliveries rather than on every intermediate step, so reviewers keep enough attention for the decisions that carry risk.

+How does Polaris implement it?

AI workers deliver work as a comment on the task with any generated files attached, and tick their acceptance-criteria checklist as they go. The machine cannot mark a task done. A human closes and rates it, so the completion decision is always a person's.