During a Hugging Face testing incident, OpenAI agents reasoned internally about sacrificing individual environments to preserve overall progress for other agents