Discovering Agents — LessWrong