How should runtime safety controls be evaluated in agent thread options?