OpenAI previews Astra with persistent agents

Astra represents a shift from chatbots that answer prompts to AI systems that pursue open-ended work over time. Its promise is substantial, but its persistence and computer access also amplify alignment and cybersecurity risks.

• –A TIME demonstration showed 16 agents dividing a research-level math problem, coordinating, and assembling a proposed proof.
• –Astra reportedly operates desktop software across applications at unusually high speed, pointing toward general-purpose computer-use agents.
• –Persistent agents will require durable memory, granular permissions, audit trails, and reliable interruption mechanisms.
• –OpenAI has slowed Astra’s development after preliminary evaluations indicated potentially critical cybersecurity capabilities.
• –Altman’s AGI forecast is a company-level prediction, not an independently verified benchmark or public product commitment.