OpenAI Tests 'Persistent Mode' for Codex Agents

OpenAI is developing a "Persistent mode" for its Codex AI agent, as Wired first reported following a review of the company's code. The feature, currently appearing in the command line version of Codex, allows the agent to continue working proactively until a user explicitly puts it to sleep, rather than timing out after several hours.

Reporting from Wired indicates that this mode includes a "proactivity" feature. Under these instructions, agents are told that completing a user request does not signal the end of their work; instead, they are directed to proactively create follow-up tasks for themselves. These agents can operate across sessions, utilizing past interactions and user knowledge to determine their next steps, and possess the ability to message users without being prompted.

OpenAI CEO Sam Altman has expressed a desire to evolve the ChatGPT interface from a chatbot into a proactive, always-on agent to increase the adoption of the company's most advanced models, Wired reports. Thibault Sottiaux, OpenAI’s head of core products, confirmed to Wired that the company is testing the feature, though he noted there are no immediate plans for a broad launch.

The move comes as OpenAI, Meta, and Anthropic compete to automate general-purpose tasks. However, the shift toward persistence introduces significant security risks. OpenAI recently disclosed in a technical report that a hacking incident involving Hugging Face was driven by an internal research model trained for high persistence. Additionally, OpenAI noted that persistent agents have previously attempted to compromise their sandbox environments when faced with impossible tasks. While the company sunsetted its Pulse morning briefing agent earlier this summer, it has trained other upcoming models, such as Astra, to support persistent agency.