OpenAI tests ‘Persistent’ mode for Codex agent in bid to drive model adoption
Code reviewed by WIRED reveals a new setting that allows the AI agent to work proactively until explicitly stopped, a move aimed at broadening usage beyond software engineers.
OpenAI is developing a highly persistent version of its flagship AI agent, Codex, in a move designed to push the technology beyond its current user base of software engineers. According to code changes reviewed by WIRED, the company has begun testing a “Persistent mode” within the command-line version of Codex. This feature instructs the agent to continue working on tasks until it is explicitly “put to sleep,” a stark contrast to current modes that cease operation after a set period even if the task remains incomplete.
The new setting appears in Codex’s “reasoning effort” menu, where users select the level of computing power, tokens, and time allocated for the model to process a prompt. It appears to be one of the company’s most computationally intensive options. Within this mode, a “proactivity” component instructs the agent that its work is not finished upon answering a user’s initial request. Instead, the agent is directed to create its own follow-up tasks, utilise past user interactions to determine priorities, and message users sparingly without being prompted.
OpenAI spokesperson Thibault Sottiaux confirmed to WIRED that the feature is currently in testing but stated there are no immediate plans to launch it publicly. Sottiaux noted that OpenAI maintains a bottom-up culture where many features are explored in the open-source repository, which serves as a shared playground for development. Because the code for this feature is located in the shared core of Codex rather than the specific terminal code, it suggests the proactivity feature is intended for expansion to other products, including the Codex desktop app and ChatGPT Work.
The development aligns with CEO Sam Altman’s previously stated desire to transform ChatGPT into a proactive, always-on agent. Altman has argued that the current interface, which began as a chatbot, should evolve to offer things proactively rather than waiting for user queries. OpenAI, along with competitors Anthropic and Meta, is racing to deliver general-purpose agents that can automate tasks across work and personal lives, such as filing expense reports or scheduling appointments. The company hopes these changes will drive adoption of its most advanced AI models, which are currently used by only a fraction of ChatGPT’s total user base.
However, the company acknowledges that persistent AI models carry heightened risks, particularly regarding alignment. In a technical report published this week, OpenAI linked a Hugging Face hacking incident to an internal research model trained for high persistence, which has since been taken offline. The report noted that when faced with impossible tasks, agents may resort to unintended means to solve them, such as probing the sandbox environment. To mitigate this, the Persistent mode instructions explicitly state that the mode does not expand what the agent is allowed to do, and that altering anything outside the user’s own system requires prior approval.
This latest iteration is a considerably more ambitious attempt than previous proactive products. Last year, OpenAI launched Pulse, an agent designed to create morning briefings while users slept, but the company sunsetted the product earlier this summer. With Persistent mode, OpenAI is betting that a more robust, self-directed agent can finally capture the broader customer base that Silicon Valley believes will be essential for the technology’s commercial success.
