OpenAI is developing a proactive, persistent version of its flagship AI agent, Codex, WIRED has learned. The new feature could allow Codex to continue working on tasks, create follow-up assignments, and operate across multiple sessions.
OpenAI recently began adding code for a new “persistent mode” setting to the command-line version of Codex. The change is part of the codebase for products reviewed by WIRED. Updates to Codex’s command-line tools are published by default, and new features often appear there before being introduced to OpenAI’s other agent products, including the Codex desktop app and ChatGPT Work.
Persistent mode has not yet been widely deployed or officially announced, and OpenAI may ultimately decide not to release it. An OpenAI spokesperson confirmed to WIRED that the company is testing the feature but said there are no immediate plans to introduce it.
“OpenAI is a very bottom-up culture, with a lot of things being considered in open source repositories that are like our shared playground,” Thibault Sautiaux, OpenAI’s head of core products, said in a statement to WIRED.
The feature appears to be OpenAI’s latest effort to develop an AI agent that people will use regularly. OpenAI, Anthropic, and Meta are competing to build general-purpose AI agents capable of automating tasks across work and personal life, from filing expense reports to scheduling doctor’s appointments. Although AI agents are currently used primarily by software engineers, Silicon Valley expects the technology to become a major business category with a much broader customer base.
Persistent mode appears in Codex’s “Inference Effort” menu, which lets users select how much computing power, how many tokens, and how much time an AI model can use to process a prompt. The setting appears to be among OpenAI’s most computationally intensive configurations. When a user selects persistent mode, OpenAI’s code indicates that Codex “continues to work until it goes to sleep.” That differs sharply from existing modes, in which Codex stops working after a few minutes or hours, even when a task remains unfinished.
Elsewhere in the Codex codebase, OpenAI describes a feature within persistent mode called “Proactive.” It appears to function as a system prompt for agents operating in persistent mode, telling them that their work is not necessarily complete after responding to a user’s request. Instead, the agents are instructed to generate their own follow-up tasks and continue working on them across sessions. They can use previous user interactions and “user knowledge” to determine what to work on next. Some tools can send unsolicited messages to users, although the instructions say those messages should be used sparingly.
The file also outlines limits for the agent. It states that persistent mode does not grant additional permissions and that the agent must receive user approval before changing anything outside the user’s own system. Those restrictions appear designed to reduce the risks associated with always-on AI agents. Because the file is located in Codex’s shared core rather than in device-specific code, it also suggests that the proactive feature could eventually extend beyond the command-line tool.
In a recent podcast, interview, and private investor meeting, OpenAI CEO Sam Altman discussed the company’s plans for ChatGPT and an proactive, always-on AI agent. The company hopes developments such as persistent Codex could accelerate adoption of its most advanced AI models, which are currently used by only a portion of ChatGPT’s overall user base.
Source: www.wired.com