Complex Behavior from Simple (Sub)Agents — LessWrong