Codex for (Almost) Everything

The boundary between a coding assistant and a computational partner is being redrawn. For the more than 3 million weekly developer users of Codex, the tool is no longer confined to the editor; it is being given control of the entire machine.

A major update to Codex introduces capabilities that fundamentally alter its operational scope. (Source). The system can now operate a computer by seeing, clicking, and typing with its own cursor. This is coupled with the ability to schedule future work for itself and automatically resume long-term tasks that may span days or weeks. To expand its reach, more than 90 additional plugins are being released, and it can now generate and iterate on images using the gpt-image-1.5 model. A new memory feature allows Codex to remember context from previous experiences, including personal preferences and corrections. These updates are rolling out to Codex desktop app users signed in with ChatGPT, with the computer use feature initially available on macOS before a planned release for EU and UK users.

The analysis is straightforward. The shift is from passive generation to active agency. Giving Codex its own cursor to see, click, and type moves the system from being an input to an application to being an operator of the application layer itself. This is a significant change in the interaction model. When combined with the ability to schedule its own work and maintain memory of user preferences, Codex is being positioned not as a tool for discrete tasks, but as a persistent agent capable of managing complex, multi-day projects.

The integration of gpt-image-1.5 and more than 90 plugins is not merely about feature expansion. It is about broadening the system’s aperture for both input and action. By processing and generating images, Codex gains a visual modality. By integrating with more tools, it gains a wider field of action within a user's digital environment. The requirement for a ChatGPT sign-in suggests a strategy to unify user context and memory across services, making the system's learning more durable and personalized. The initial limitation to macOS is a tactical deployment choice, but it establishes a clear path for how this agency will be introduced to other platforms.

This update leaves one critical question on the table. As Codex transitions from a tool that is wielded to an agent that is delegated to, what is the appropriate model for supervision and control? When a system can operate a machine autonomously over weeks, remember past corrections, and interact with any application via the graphical interface, the nature of oversight changes. The core challenge is no longer just about the quality of a single output, but the trajectory of a long-running process.