OpenAI brings Sky technology into Codex for Mac and strengthens background app control

OpenAI has turned Codex for Mac from a code-writing tool into a full desktop operator: the agent can now control apps in the background, switch between several windows and perform routine actions without pulling them to the foreground and without requiring constant user intervention. The key update comes from Sky, a macOS startup that OpenAI officially acquired together with its development team. At the product level, this no longer looks like a side experiment.

Codex now has a built-in computer-control mode that lets the agent see where it clicks, type text, open apps and work in several windows at once. Importantly, all of this happens in the background: the user can keep working while the agent handles tasks in other programs in parallel. For the Mac environment, where many processes are still poorly automated, this is a notable shift.

The new approach is technically interesting because Codex does not rely only on screen recording and simulated clicks on an image. Instead, the system uses macOS accessibility mechanisms, which can read the hierarchy of interface elements: buttons, input fields, lists, tabs and other objects inside a window. For a language model this is far richer than an ordinary screenshot: it receives not just a picture but a structured description of the interface, and can better understand what is in front of it. As a result, the agent finds it easier to choose the right action and depends less on brittle scenarios built on cursor coordinates or workaround scripts.

That is why the new version of Codex looks stronger than many competing solutions. Most computer-control systems today are still built on a combination of screenshots, OCR and basic automation through system commands. That approach works as long as the interface is predictable, but it breaks quickly in real applications. According to early reviews, Codex copes better with tasks that require navigating complex windows, scrolling long feeds, opening nested elements and executing chains of actions in sequence. Parallel cursors are a separate advantage: several agents can work on one Mac at the same time without getting in each other's way and without stealing focus.

The practical value of the update shows in everyday scenarios. Codex can open a music app and play the right album, even if the command is not phrased perfectly. It can look through several work programs, such as a messenger, an RSS client and a reader, and then compile a short summary from them. It can install files, open them, check the result and continue debugging without manual switching between windows. A person often does these things faster, but that is not the main point: the agent takes over routine work that can run in the background, especially where there is no API, no console commands and no full AppleScript support.

For OpenAI this is also a strategic step. Codex looks less and less like an ordinary coding assistant and more and more like a control center for agents. The company first released the Codex app for macOS on 2 February 2026, and on 16 April added a major update with background computer control, memory, automations and new plugins. According to OpenAI, the service is already used by more than 3 million developers per week, and the company is clearly trying to extend its use cases beyond the IDE: from writing code to testing interfaces and working with documents, the browser and desktop apps.

The main takeaway is simple: OpenAI is betting not only on stronger models but also on deeper control over the user's environment. If Codex really maintains its current accuracy and speed, the AI agent market could quickly shift from the chat window to background assistants that work in real programs as confidently as they work with text and code today.