OpenAI's Astra agents run for days — but error compounding remains unsolved
openai is testing 'astra', a multi-agent model family designed to coordinate tasks for hours or days, marking its first run through a new us government review framework.
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.