Code changes to the Codex command-line tools are publicly available by default, and new OpenAI features typically first appear there before being extended to products like the Codex desktop application and ChatGPT Work.
By Frank Ulom
·
Published on August 28, 2026
·
3 min read
According to a report by Wired magazine on August 27 local time, OpenAI is developing Codex, a more proactive AI agent capable of working continuously for extended periods, hoping to free its flagship AI agents from simply waiting for user commands.
Wired’s review of recent code changes to Codex revealed that OpenAI has recently begun adding a “persistent mode” to the Codex command-line version . Code changes to the Codex command-line tools are publicly available by default, and new OpenAI features typically first appear there before being extended to products like the Codex desktop application and ChatGPT Work.
Persistent mode is currently in the testing phase and has not yet been widely rolled out or officially released. An OpenAI spokesperson has confirmed that the company is indeed testing the feature, but has no plans to launch it in the near future.
Thibault Sotio, OpenAI’s core product lead, said: “OpenAI’s culture is very bottom-up. People will try out a lot of different things in the open-source codebase. It’s a bit like our common testing ground.”
Persistent mode appears in Codex’s “Inference Strength” menu, allowing users to decide how much computing power, tokens, and time to allocate before the model answers a question .
Meanwhile, persistent mode is currently one of OpenAI’s most resource-intensive settings. When enabled, Codex will continue to work until it is put into a dormant state . Other modes typically stop running after a few minutes or hours, even if the task is not completed.
OpenAI has also designed a capability called “proactivity” for persistent mode. Based on system prompts in the code, the agent cannot consider the task finished after answering the user’s question; instead, it needs to proactively schedule follow-up work for itself . The agent can continue performing these tasks across sessions and, based on past interactions with the user and “understanding the user,” independently determine what to do next.
The system even supports agents sending messages proactively without a request from the user, but this should be used sparingly.
OpenAI also defines the permission boundaries for this persistent agent. Persistent mode does not expand the agent’s original operational permissions; if modification of content outside the user’s own system is required, user approval must be obtained first .
The code containing the “proactive” feature is not a dedicated part of the Codex terminal tool, but rather resides in the Codex shared core. The report suggests this indicates OpenAI may plan to bring this capability to more products.
Daily Lamp reports that, OpenAI CEO Sam Altman has recently stated publicly on several occasions that he hopes to eventually transform ChatGPT into a proactive and continuously operating AI agent . OpenAI also hopes to use this to increase the usage rate of its state-of-the-art models, as currently only a small portion of ChatGPT’s large user base is actually using these models.
OpenAI has made several attempts to develop proactive AI products, but has yet to truly penetrate the market. For example, Pulse, launched last year, generated morning briefings while users slept, but was discontinued in early summer this year.
