OpenAI is developing a new version of its AI agent, Codex, with a feature called Persistent mode. This mode allows Codex to continue working on a task until it's completed, making it more akin to a personal assistant that never sleeps. Users can select the level of computing power needed for Codex to 'think' before answering a prompt.

In Persistent mode, Codex is instructed to proactively create follow-up tasks for itself, using past user interactions and 'knowledge of the user' to decide what to work on. However, the agent is limited in its actions, requiring user approval to make any changes outside the user's own system.

The feature represents OpenAI’s latest attempt to make AI agents more useful for people’s daily lives. OpenAI, Anthropic and Meta are competing to deliver general-purpose agent products that could help with tasks such as filing expense reports or scheduling appointments. Currently, the people who use AI agents are mainly software engineers, but the tech is seen as a potential major line of business with a broader customer base.

While Persistent mode could enhance the capabilities of AI agents, it also raises concerns about user privacy and the potential misuse of the technology. OpenAI acknowledges that persistent AI models carry heightened risks, citing an internal hacking incident involving a highly persistent model. The company has since taken this model offline and is training other forthcoming AI models, including Astra, to enable persistent agents.