OpenAI Is Developing a ‘Persistent’ AI Agent
OpenAI is testing a “Persistent mode” for its Codex command-line tool that could allow an AI agent to keep working until it is explicitly put to sleep. The feature matters because it points towards a more proactive, always-on form of AI assistant, as major technology companies compete to make general-purpose agents useful beyond software engineering.
Code reviewed by WIRED indicates that the mode would use substantial computing resources and could create its own follow-up tasks across sessions, drawing on previous interactions and knowledge of the user. It could also message users without being prompted, although instructions say this should be done sparingly and that it cannot take actions outside a user’s system without approval; OpenAI confirmed testing but said it has no immediate launch plans.
- OpenAI is testing an always-on Codex agent mode.
- It could continue tasks and create follow-up work.
- OpenAI says no immediate launch is planned.
New here? Start with this
AI agents are software designed to carry out tasks with some independence, rather than only responding to each individual prompt. Codex is OpenAI’s tool for working with computer code, and a persistent setting could let an agent continue its work across separate sessions.
A more proactive assistant might remember earlier interactions, suggest follow-up tasks or contact a user without being asked. That could make agents useful for more than coding, while raising questions about computing costs, user control and when an assistant should act on its own.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
A persistent agent could make AI more useful by carrying work across sessions, remembering relevant context and following up without requiring users to restart or manage every step. If it uses computing resources responsibly and seeks approval before acting beyond the user’s system, proactive assistance could save time while leaving important decisions under human control.
The case against
An agent that runs continuously and draws on past interactions may consume substantial computing resources and raise concerns about privacy, oversight and unwanted messages. Even with limits on external actions, users may find it difficult to understand or control what the agent is doing, so caution about developing or deploying such a system can be justified.