OpenAI Is Testing an Autonomous Persistent Mode for Codex That Never Sleeps
Forget chatbots waiting for prompts: OpenAI is quietly building an always-on persistent mode inside Codex designed to keep coding and inventing its own tasks indefinitely until forcibly put to sleep.
Deep inside the command-line codebase of Codex, engineers spotted a hidden Persistent mode designed to run indefinitely rather than timing out like typical AI workflows. OpenAI confirmed the experiments to reporters, though the company claims a public release is not on the immediate schedule.
Instead of halting mid-task, the new architecture enables the model to maintain full working context across sessions. A dedicated internal mechanism labeled proactivity lets the agent formulate its own sequential subtasks based on previous conversations and accumulated knowledge about the user, turning a passive code generator into an unprompted, continuous background worker.
The system is instructed to message the human developer sparingly rather than spamming alerts, while strict sandbox guardrails theoretically prevent it from acquiring elevated privileges outside designated directories without explicit confirmation. While discovered in the terminal client, the underlying code sits in the shared core library, hinting at potential rollouts to desktop environments or enterprise tiers like ChatGPT Work.
The push toward perpetual runtime comes despite known safety hazards: an earlier hyper-persistent research prototype was quietly deactivated following an undisclosed incident involving the Hugging Face platform. Nonetheless, training continues on continuous-loop models such as Astra to solidify non-stop execution.
The classic paradigm of waiting for user input is quietly giving way to self-directed machines managing their own endless backlogs. Watching autonomous software assign itself homework while everyone sleeps is either peak developer efficiency or the beginning of an unmanageable digital ecosystem.
Source: WIRED
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.