OpenAI 的 Astra 智能体可运行数天 — 但错误累积问题仍未解决
openai 正在测试 'astra'，这是一个旨在协调持续数小时或数天任务的多智能体模型家族，标志着其首次通过新的美国政府审查框架。
openai is building astra, a model family designed to handle tasks lasting hours or days by coordinating multiple agents, with ceo sam altman showcasing the tech to us regulators. technical details reveal a potential launch as gpt-6 or a gpt-5 variant, aiming to achieve autonomous research capabilities by 2028. the industry must solve the issue of error accumulation in long-horizon workflows, where coordination overhead often degrades performance on complex planning tasks.