OpenAI Builds Persistent AI Agent for Continuous Tasks

OpenAI is currently engineering a persistent AI agent designed to handle continuous, long-running tasks without requiring constant user intervention. According to internal code reviewed by WIRED, the company has begun integrating a new “Persistent mode” setting into the command-line version of its flagship coding assistant, Codex. Because updates to the open-source command line tool are publicly visible and often precede features in consumer-facing products like the ChatGPT desktop app or ChatGPT Work, this development signals a significant shift in how OpenAI intends to deploy its most advanced models.

While the feature remains in the testing phase and has not been officially launched, an OpenAI spokesperson acknowledged that the company is actively exploring the capability. “OpenAI is a very bottoms up culture and many different things are explored on the open source repo which is a bit of our shared playground,” said Thibault Sottiaux, OpenAI’s head of core products, explaining the experimental nature of the codebase. The company stated there are no immediate plans to release Persistent mode broadly, but it represents a strategic move toward creating AI agents that users find genuinely useful for everyday automation.

Beyond Simple Command Response

The new setting appears within Codex’s “reasoning effort” menu, allowing users to allocate specific levels of computational power, tokens, and time for the model to process complex prompts. This configuration is described as one of the most resource-intensive options available. Unlike standard modes that terminate a task after a few minutes or hours regardless of completion status, the internal code indicates that in Persistent mode, Codex will “continue working until put to sleep.” This distinction aims to solve a major pain point in current AI interactions: the inability of agents to complete multi-step workflows autonomously.

Silicon Valley giants including Anthropic and Meta are similarly racing to develop general-purpose agents capable of automating routine personal and professional duties, such as scheduling appointments or managing expense reports. Although software engineers currently dominate the user base for these tools, industry leaders believe persistent, proactive agents could unlock a much broader market. By enabling the AI to work across sessions and leverage past interactions, OpenAI hopes to transform ChatGPT from a reactive chatbot into a truly helpful digital assistant.

Proactive Behavior and Safety Constraints

A key component of this new architecture is what the code describes as “proactivity.” Agents operating in Persistent mode receive system instructions telling them that their job is not finished once they answer an initial query. Instead, they are directed to generate follow-up tasks for themselves based on their “knowledge of the user” and previous context. The agent also possesses the ability to message the user without being prompted, though the instructions emphasize sending such messages sparingly to avoid annoyance.

Despite these expansive capabilities, the code includes strict safety boundaries. The documentation explicitly states that Persistent mode does not expand the scope of actions the agent is permitted to take. Any attempt to alter systems outside the user’s own environment requires explicit prior approval. These constraints appear to be located in the shared core of Codex rather than terminal-specific code, suggesting that the proactivity framework is intended to support multiple interfaces beyond just the command line.

This push for always-on intelligence aligns with recent statements from OpenAI CEO Sam Altman, who has expressed a desire to turn ChatGPT into a proactive, always-available agent. The ultimate goal is to drive higher adoption of OpenAI’s most sophisticated models, which currently serve only a fraction of the platform’s total user base. By offering agents that can persistently manage complex projects, OpenAI aims to demonstrate tangible value that encourages wider usage among non-technical consumers.