AI Agent Security (Course)
Approval Gate
Also: tool-call gate
The point in an agent loop where a proposed tool call is allowed, denied or sent to a human before it runs.
In more detail
The gate is the only place a rule can actually be enforced, because it sits between the model's request and the real action. It may be a prompt to the user, a permission rule, a hook that exits with a blocking status, or a sandbox that makes the action impossible. A gate that depends on the model's own judgment is weaker than one that runs as ordinary code, and a gate that matches command text can be bypassed by rephrasing the command.