Human-in-the-loop without the bottleneck

"Human-in-the-loop" is supposed to be a safety feature. Done wrong, it becomes the reason the agent is useless: if a person has to approve every call, you have not automated anything, you have added a queue. The skill is choosing which moments deserve a human.

The bottleneck is a design choice

Approving everything trains people to approve without reading — and a reflexive click is worse than no gate, because it manufactures the appearance of oversight. The fix is selectivity: gate the few actions where human judgment changes the outcome, and let the rest flow.

A workable split

• Auto: reads, analysis, drafts, dry-runs — anything cheap to undo.
• Gate: spend, publish, push, delete, customer-facing sends — anything expensive or irreversible.
• Never: actions outside the worker's role — those should not be reachable at all, not merely gated.

Make approvals fast and contextual

A good gate hands the reviewer a concrete diff — worker, action, account, exact payload — so the decision takes seconds, not a research session. The agent pauses, the human decides, the call resumes.

Oversight that fires on everything is noise. Oversight that fires on what matters is control.

Grantry lets you mark which actions are risky per role, so humans land on the decisions worth their attention and the agent owns the rest.