OpenAI is quietly testing a persistent, always-on version of its Codex coding agent

Code spotted in OpenAI’s Codex CLI repository reveals a “Persistent mode” that lets the agent keep working until a user explicitly “puts it to sleep,” rather than timing out after minutes or hours. A related “Proactivity” feature would let Codex generate its own follow-up tasks and act on them across sessions, drawing on prior interactions to decide what to work on next. OpenAI hasn’t announced the feature or given a rollout timeline, and says any actions outside a user’s own system still require approval.

Always-on agents raise the stakes on governance basics - scoped credentials, spend limits, and a kill switch matter more once an agent no longer stops when you look away.