That's a decent take. But how do we account for people that are not most effectively modeled as agents with goals? Deontologists for example, can be evil even if their (alleged) preferences entirely match mine.