An approval gate is narrower than keeping a human in the loop. Human in the loop is the design idea. A gate is the specific place where it's built, on one named action, with one person who can clear it.
What makes a gate work
It sits on an action that's expensive to reverse, not on every action
The approver sees what the agent is about to do, in plain words
Saying no is as easy as saying yes, and the agent handles no without breaking
Both the ask and the answer end up in the record, with who answered
Where teams get it wrong
They put a gate on everything. A person who gets forty approval prompts a day stops reading them by the tenth. Now the gate is a rubber stamp that also produces a record saying a human approved it, which is worse than having no gate at all, because the record looks like evidence.
Here's a concrete one. An agent handles support tickets. Reading a ticket and drafting a reply are both cheap to undo, so neither one needs a gate. Issuing a refund isn't cheap to undo. One gate, on the refund. The reviewer gets a handful of prompts a week and actually reads them.
Count the gates you have. If the number is high enough that nobody could read them all, you don't have controls. You have a queue.