OpenAI Is Developing a ‘Persistent’ AI Agent

Details

WIRED reviewed code changes in OpenAI's public Codex CLI repository and found 'Persistent mode' actively in development. Unlike current modes that stop after minutes or hours even mid-task, this setting keeps Codex running indefinitely — 'until put to sleep.' OpenAI confirmed it's testing the feature but has no immediate launch plans. A built-in 'proactivity' system makes persistent agents genuinely autonomous: they generate their own follow-up tasks after completing a request, carry work across sessions, draw on past user interactions to decide what to work on next, and can message users without being prompted (though told to use this capability sparingly). The mode appears in Codex's reasoning effort menu alongside standard settings and is flagged as computationally intensive. Explicit guardrails are coded in: Persistent mode does not expand the agent's permissions, and any action outside the user's own system requires explicit approval. Thibault Sottiaux, OpenAI's head of core products, described the Codex CLI as 'a shared playground' for bottom-up exploration — signaling formal release could be months away. The broader stakes: OpenAI, Anthropic, and Meta are racing to grow AI agent adoption beyond today's software-engineer-heavy audience. Persistent, self-directed agents are the core bet — hand off a long-running task and have it complete autonomously, which is what's needed to automate expense reports, scheduling, or research for a far wider workforce.