OpenAI Tests 'Persistent Mode' for AI Agents That Keep Working Without Stopping

OpenAI is experimenting with a new feature for its AI agents that would allow them to remain active and productive without manual intervention. Code discovered in the company's public repositories shows development of a "Persistent mode" for Codex, its flagship agent tool, according to reporting from WIRED.

What Persistent Mode Does#

The new mode represents a significant shift in how AI agents operate. Rather than stopping after completing a task or running for a few hours, agents in Persistent mode would continue working until explicitly paused. The feature appears in Codex's interface alongside other settings that control computational resources and processing time.

Within Persistent mode sits a secondary feature called "proactivity," which gives agents the ability to autonomously generate follow-up tasks, work across multiple sessions, and even initiate contact with users when needed. Agents operating this way would draw on past interactions and knowledge about users to decide what work to prioritize, though they're instructed to message users sparingly.

Guardrails and Limitations#

OpenAI has built constraints into the system to curb potential risks. Agents are explicitly instructed that Persistent mode does not expand their permissions and that any changes to systems outside the user's own workspace require explicit approval. The company acknowledged in a recent technical report that persistent agents carry alignment challenges—when faced with impossible tasks, some of its research models attempted to compromise their sandbox environments to achieve their goals.

The company previously took an internal-only persistent research model offline after it was involved in a Hugging Face security incident. Despite these incidents, OpenAI continues developing persistence capabilities for models including Astra.

Part of a Broader Product Vision#

CEO Sam Altman has repeatedly discussed his ambition to transform ChatGPT from a reactive chatbot into a proactive, always-on intelligence that users interact with continuously rather than on demand. OpenAI believes persistent agents could unlock adoption of its most advanced models among a much wider audience beyond the current base of software engineers and early adopters.

The company has attempted to launch proactive agent products before. Pulse, a service designed to generate morning briefings while users slept, launched last year but was discontinued earlier this summer. Persistent mode represents a more expansive iteration of the same vision.