OpenAI is testing a more persistent version of Codex, its AI coding agent, according to WIRED, which reviewed recent code changes. The feature, called “Persistent mode” in the code WIRED reviewed, has not been announced or broadly rolled out. An OpenAI spokesperson confirmed to WIRED that the company is testing the feature, while saying there are no immediate plans to launch it. Thibault Sottiaux, OpenAI’s head of core products, told WIRED that OpenAI explores many ideas in its open-source repository, describing it as a shared playground for the company’s product work. The code changes appeared in the command-line version of Codex, WIRED reports. The outlet says changes to that tool are public by default and that new features often show up there before they reach other OpenAI agent products, including the Codex desktop app and what WIRED identifies as ChatGPT Work. That makes the code notable, but not equivalent to a product announcement. The central change is duration and autonomy. According to WIRED’s review, Persistent mode appears in Codex’s “reasoning effort” menu, where users choose how much compute, token budget and time to give a model before it responds. WIRED reports that the setting appears to be among OpenAI’s more computationally intensive options, and that code text says Codex would “continue working until put to sleep.” That differs from the agent behavior WIRED describes as currently available, where Codex stops after a period of work even when a task is unfinished. If the feature ships, it could move Codex further from a request-response coding assistant toward a longer-running agent that can keep pursuing work across more time. WIRED also found references to a “proactivity” capability tied to Persistent mode. The outlet reports that the code appears to instruct the agent that its job is not necessarily finished after answering a user request. Instead, it may create follow-up tasks for itself, work on them across sessions, use past user interactions and “knowledge of the user” to decide what to do, and message the user without being prompted, though sparingly. The same reviewed file also appears to set limits. WIRED reports that the instructions say Persistent mode does not expand what the agent is permitted to do, and that the agent needs user approval before changing anything outside the user’s own system. WIRED notes that the file sits in shared Codex core code rather than code specific to the terminal, which may indicate the capability is being considered beyond the command-line tool. The competitive backdrop is the agent race. WIRED frames the feature as part of a broader push by OpenAI, Anthropic and Meta to build general-purpose agents that can automate work and personal tasks. For now, WIRED reports that agent usage remains concentrated among software engineers, even as Silicon Valley sees broader commercial potential. Who benefits: OpenAI could benefit if Persistent mode turns Codex into a more useful workflow tool for developers. Early users of coding agents may also benefit if longer-running work reduces the need to restart or re-prompt unfinished tasks. Who's exposed: Users and organizations experimenting with agentic coding tools are exposed to the supervision problem: a more persistent agent may require clearer approval rules and monitoring. Competing agent developers, including the companies WIRED names in the broader race, face pressure if OpenAI turns the tested feature into a usable product.