Protecting agent boundaries — LessWrong