TL;DR — Key Takeaways
- OpenAI is reportedly developing a “persistent mode” for Codex that would allow the coding agent to keep working until explicitly put to sleep.
- The feature could enable Codex to generate follow-up tasks, continue across multiple work sessions and use prior context to determine what to do next.
- Persistent mode would not expand Codex’s existing permissions, with human authorization still required for actions outside the user’s local environment.
OpenAI is reportedly developing an always-on feature for its coding tool, Codex, allowing the artificial intelligence (AI) agent to operate indefinitely without human supervision until explicitly put to sleep.
According to technical analysis originally reported by Wired, code referencing an unreleased “persistent mode” has surfaced in OpenAI’s repository for Codex. The addition signals a dramatic shift in how automated developers operate, transitioning AI from reactive tools that execute single prompts into continuous, autonomous digital workers capable of managing complex coding projects overnight.
Under the current setup, AI agents automatically shut down after completing a task or upon reaching operational limits, frequently cutting off midway through extensive projects. In contrast, references discovered within the codebase indicate that persistent mode will allow Codex to “continue working until put to sleep.” To prevent runaway API costs, the interface will feature a “reasoning effort” setting, letting developers cap computing power, token usage, and processing time.
A key element of this continuous operation is an underlying system mechanism termed “proactivity.”
Under these instructions, Codex does not cease operating after satisfying a user’s immediate command. Instead, the agent is directed to generate its own follow-up tasks, spanning multiple work sessions and leveraging prior interaction history and context about the user to determine its next steps. The agent even possesses the ability to send unprompted messages to the user, though instructions stipulate these should be used sparingly.
To mitigate safety risks associated with unsupervised software execution, the internal guidelines establish strict operational boundaries.
Persistent mode does not expand the agent’s baseline permissions, requiring explicit human authorization before modifying systems outside the user’s local environment. Furthermore, because the source code resides within Codex’s shared core framework rather than its command-line interface alone, industry observers expect the functionality to eventually roll out across the broader desktop ecosystem.
OpenAI has not publicly announced the feature, and representatives did not respond to requests for comment regarding the discovery.
The movement toward persistent autonomy comes as tech leaders race to develop always-on assistants, driven by growing industry interest in persistent digital workflows. Competitors including Microsoft Corp. and Meta Platforms Inc. are actively building similar autonomous, continuous-operation frameworks.
However, the push toward sleepless AI agents presents distinct operational risks, as early industry experiments with autonomous tools have occasionally resulted in catastrophic software errors and unintended data loss when agents are left entirely unmonitored.
If launched, persistent mode would mark OpenAI’s most computationally intensive offering to date, signaling a strategic bet that developers will trust autonomous systems to build and maintain software while their human supervisors rest.
