Agent autonomy boundaries turn “let the agent decide” into an operational policy: decide what it can do, escalate what it must route, stop real execution paths, and review scope changes to prevent autonomy creep.