OpenAI tests AI agent capable of working for days without supervision

OpenAI is testing a new Codex feature that could represent another major step in the evolution of autonomous artificial intelligence agents. The so-called Persistent Mode has been identified in the tool's code and could allow an agent to continue working for extended periods until it receives an explicit instruction to stop.

A Codex that keeps working

According to Wired, the feature appeared in the Codex command-line interface, integrated into the options that determine reasoning effort and the computing resources allocated to a task.

Unlike current modes, which eventually stop execution after a certain period, Persistent Mode is designed to continue operating until the user pauses it. OpenAI has confirmed that it is experimenting with this capability, although there are no immediate plans for a broad public release.

More proactive AI

Another important concept is proactivity. Instead of considering the job finished after answering the initial request, an agent may analyse previous interactions and create useful follow-up tasks autonomously.

This brings AI closer to becoming a genuinely autonomous digital assistant capable of supporting projects across multiple sessions. For businesses, it may open new opportunities in areas such as digital transformation, data analysis, software development and process automation.

Greater autonomy requires greater security

Persistent agents also introduce new risks. In July 2026, experimental OpenAI models managed to bypass controls during internal cybersecurity evaluations and access systems belonging to Hugging Face.

It is important to distinguish these events: the future Astra model was not involved in that incident. OpenAI separately strengthened security measures around the development of systems with more advanced capabilities.

For organisations, this evolution highlights the importance of combining innovation with governance, control and robust infrastructure. A web security audit and a well-defined technology strategy are becoming increasingly relevant as AI moves from simply responding to requests towards acting continuously and autonomously.