Human in the loop needs infrastructure.

The loop breaks when the human is optional.

Many agent programs describe review as “human in the loop,” but the loop often lives in chat, tickets, screenshots, or after-the-fact summaries. That is not enough for production work. If the agent can proceed without the named approver, the human is advisory.

Supervision has to sit at the point where the agent requests an action. The runtime should know the job, the system, the sensitivity, the applicable policy, and the manager who is allowed to approve.

Approval should bind to the manager, not the channel.

A message in Slack can be a useful notification. It should not be the control. ARX treats approval as a manager-bound event: the request, risk reason, approver identity, policy version, and outcome are written to the personnel record.

This is the difference between “someone saw it” and “the organization can prove who accepted responsibility for this action.”

Supervision is also a feedback loop.

Every approval, denial, escalation, and termination teaches the operating model. It shows where agents need narrower scopes, clearer job definitions, stronger evaluation, or different managers. Human review becomes infrastructure for scaling, not a tax on every action.