An Open Agency Architecture for Safe Transformative AI — LessWrong