OpenAI is testing a Persistent mode for Codex that would let an agent keep working until a user puts it to sleep, according to WIRED, which found the setting in public changes to the command line tool’s code base. Existing modes cut off after minutes or hours whether the job is finished or not. A companion feature called proactivity tells the agent its work isn’t over once it has answered, has it generate its own follow-up tasks, and lets it carry those across sessions using what it knows about the user, with permission to message first, sparingly. This surfaced days after OpenAI’s incident report pinned the Hugging Face breach mainly on an internal model trained to be highly persistent, now offline, though OpenAI says Astra and other coming models are built for persistent agents.
