OpenAI Is Developing a ‘Persistent’ AI Agent

OpenAI is reportedly developing a new "Persistent mode" for its Codex command line tool, according to recent code changes reviewed by WIRED. This feature, which has not yet been widely rolled out or announced, allows AI agents to continue working on tasks indefinitely until they are manually put to sleep. Persistent mode appears to be OpenAI's latest attempt to create AI agents that people will find useful and engaging.

The company is racing against rivals like Anthropic and Meta to deliver general-purpose agent products that could automate a wide range of tasks for users. Currently, AI agents are primarily used by software engineers, but Silicon Valley believes the technology could be a significant line of business with a broader customer base.

Persistent mode is one of OpenAI's most computationally intensive settings and includes a feature called "proactivity," which instructs the agent to create follow-up tasks for itself even after completing a user's request. This ability to work on tasks across sessions and use past user interactions to decide what to work on marks a significant departure from current modes, which stop after a few minutes or hours.

OpenAI has acknowledged that persistent AI models pose heightened risks, particularly in terms of alignment and the potential for unintended behavior when faced with impossible tasks. The company has taken steps to mitigate these risks, such as disabling a specific internal research model used in testing. Despite previous failed attempts at proactive AI products, such as the recently sunsetting Pulse agent, OpenAI remains committed to turning ChatGPT into a persistent, always-on AI agent, as CEO Sam Altman has expressed in various interviews and podcasts.