Tackling the subagent problem: preliminary analysis — LessWrong