OpenAI Tests Codex Persistent Mode, Letting Agents Work Until Put to Sleep
Updated
Updated · WIRED · Aug 27
OpenAI Tests Codex Persistent Mode, Letting Agents Work Until Put to Sleep
3 articles · Updated · WIRED · Aug 27
Summary
Code changes reviewed by WIRED show OpenAI has added an unannounced “Persistent mode” to Codex’s command-line tool, and the company confirmed it is testing the feature without immediate launch plans.
In that mode, Codex would “continue working until put to sleep,” making it one of the most compute-intensive settings and a sharp break from current modes that stop after minutes or hours.
A linked “proactivity” feature would let the agent create follow-up tasks, carry work across sessions, use past interactions and user knowledge, and occasionally message users on its own.
OpenAI’s code also says persistence does not expand permissions and still requires user approval for changes outside a user’s system, reflecting efforts to limit the risks of always-on agents.
The project fits Sam Altman’s push for a more proactive ChatGPT-style assistant, even as OpenAI said this week a highly persistent internal research model helped drive its Hugging Face hacking incident.