OpenAI is developing a 'persistent' AI agent

OpenAI is developing a new capability for its Codex system that could fundamentally change how AI agents operate. According to code reviewed by WIRED, the company is building a 'persistent' agent feature that allows Codex to continue working on tasks proactively until it is explicitly 'put to sleep.' This marks a significant shift from the current model, where AI agents only act when prompted by a user. The concept is simple but powerful. Instead of asking Codex to write a piece of code and then waiting for the result, users could assign a larger goal—such as 'refactor the entire authentication module'—and Codex would work on it autonomously, making decisions, writing code, running tests, and fixing issues until the task is complete. The agent would only stop when the user tells it to, or when it determines the task is finished. This persistent agent capability represents a move toward true autonomy in AI systems. It is a departure from the reactive paradigm that has dominated AI development, where every action requires a trigger. With persistence, AI agents can manage long-running projects, monitor systems, and respond to changes in real time, all without human intervention. The implications for software development are enormous. Developers could delegate entire feature sets to Codex and focus on higher-level architecture and design. For businesses, this could mean faster development cycles and reduced costs. But it also raises questions about oversight. If an AI agent is working autonomously, who is responsible for its mistakes? How do you ensure it aligns with the user's intent over a long period? OpenAI has not officially announced the feature, and it is unclear when it will be released. However, the code review suggests that the company is serious about pushing the boundaries of agentic AI. A persistent Codex would not just be a tool; it would be a colleague. The era of AI agents that work while you sleep is closer than we think.