The "loop" is the sense-decide-act cycle that any automated process runs. Human-in-the-loop inserts a person into that cycle as a gate: the system prepares an action, hands it over, and waits. Human-on-the-loop is the weaker cousin — the system acts on its own while a person watches and can pull a brake after something has already gone out. The difference matters most for irreversible actions: a person monitoring can stop a bad email run partway through, while a person at a checkpoint stops it before the first message leaves.
Good HITL design puts checkpoints where the stakes are, not everywhere. Requiring sign-off on every internal step creates approval fatigue and quietly trains people to click yes without reading; requiring none removes the safety entirely. The practical line runs between internal work — researching, scoring leads, drafting — which can proceed freely, and outward-facing actions — sending, publishing, spending — which need explicit sign-off. The checkpoint also has to be enforced where the action actually executes: a review screen the automation can route around is theater, not HITL.
Brohns builds its agent teams on exactly this split. Agents find leads, qualify them, and draft outreach on their own, but every email, SMS, post, or spend decision lands in the Approvals queue first, and the gate is enforced server-side — the recipient and content of an approved message come from the database, never from an agent improvising at send time. The loop is also adjustable: you start by approving everything, and as an agent earns trust you can grant it routine autonomy within limits, deliberately and reversibly shifting from human-in-the-loop toward human-on-the-loop.