This is an interesting idea.. do you use agents with more adversarial / cynical prompts whose only job is approving api calls? "You are an API approval agent on the lookout for suspicious or unusual stripe api calls...."