You can use the policy agent to define rules that reject or reshape responses that contain disallowed words or that would cause excessive token usage.