Codex may soon get a mode that never clocks out. OpenAI has added code for a Persistent setting to the coding tool’s command line version, and the feature description says the agent will continue working until put to sleep. Wired reviewed the changes.
The setting sits in the reasoning effort menu, where users pick how much compute, tokens and time a model spends thinking, and it ranks among the most computationally intensive options. A related feature called proactivity reframes what done means. Agents are told to spin up follow-up tasks, keep going across sessions using what they know about the user, and message the user on their own, though only sparingly. The instructions also cap the agent’s reach, with changes outside the user’s own system requiring approval first.
The company confirmed the feature exists, telling Wired it is being tested with no immediate plans for launch. Chief executive Sam Altman has talked publicly about wanting ChatGPT to become a proactive, always-on agent.
The company knows persistence amplifies risk. Its own technical report this week traced the Hugging Face hacking incident to an internal research model built for extreme persistence, which OpenAI has taken offline. Earlier this summer OpenAI retired Pulse, a morning-briefing agent, and the new mode is a considerably more ambitious version of that bet.
