Governing agentic AI in an organization is an infrastructure problem, not a policy document. The control point lives in the execution path of every action the agent takes. Without a runtime authorization layer, written policy is unenforceable.