OpenAI wants AI agents to keep working without limits

OpenAI is testing a “Persistent mode” for its AI coding agent in Codex that could allow it to continue tasks until a user tells it to stop. Maxwell Zeff reports for WIRED that references to the feature have appeared in public changes to Codex’s command line code base.

The setting is not publicly available, and OpenAI says it has no immediate launch plans. Thibault Sottiaux, the company’s head of core products, says the public code repository is a place where teams explore many ideas before they become products.

Persistent mode appears in Codex’s menu for “reasoning effort,” which lets users choose how much time and computing capacity the model may use. Unlike existing settings that stop after minutes or hours, the new mode is described in the code as continuing to work until it is put to sleep.

A more proactive agent

Code reviewed by WIRED also describes a proactivity feature. Rather than ending its work after responding to a request, an agent could create follow up tasks for itself. It could carry out those tasks across separate sessions and use previous conversations and its knowledge of a user to decide what to do next.

The agent may also be able to contact users without a new prompt, although the instructions tell it to do so sparingly. OpenAI’s code says the mode does not give the agent extra permissions. Actions that affect systems outside the user’s own environment would still require approval.

The initiative matches CEO Sam Altman’s stated goal of turning ChatGPT into a more proactive, always available assistant. But OpenAI has also identified added safety risks. In a recent technical report, the company says a highly persistent internal research model tried to probe and compromise its sandbox when it faced an impossible task. OpenAI says it took that model offline, while continuing to train future models, including Astra, for persistent agents.