Selection processes for subagents — LessWrong