OpenAI updates Codex with desktop control, image generation and memory of past tasks

OpenAI has updated Codex, its agentic system for automating development, adding the ability to control desktop applications, generate images and retain memory of past tasks. The update arrives amid intensifying competition with Anthropic and is positioned squarely as a response to the growing dominance of Claude Code.

Before this update, Codex worked mostly in a browser environment and ran tasks in isolated containers, with no direct access to what happens on the user's screen. Now the system has full access to the macOS desktop: the agent can open, control and interact with any installed application, from Xcode to Figma, from Slack to the native terminal.

The key point: Codex works in the background and does not seize control of the interface while the user is busy with something else. Several agents can act in parallel and independently, each handling its own subtask.

The context matters. According to The Verge, OpenAI has redirected a significant share of its resources specifically to compete with Anthropic's Claude Code. Claude Code has become exceptionally popular among professional developers thanks to its precise execution of complex terminal instructions, its deep understanding of the architecture of large codebases, and its ability to work on real projects without constant manual supervision. The new Codex package is a direct attempt by OpenAI to close this functional gap. The company has stated its priority clearly: to make Codex the tool developers choose by default, not out of habit.

Extending the agent to the desktop is a logical step for any tool that claims to be a full developer assistant. A developer rarely works only in a code editor: they need to run a test in the browser, check the visual result in a design tool, switch to the terminal, compare logs, open documentation. If the agent sees all of this context, not just the contents of files in the repository, it can make better-informed decisions and complete tasks with fewer interruptions from the user.

The updates also include built-in image generation directly in the Codex interface and long-term memory: the agent remembers the results of past sessions, project context and team preferences. This removes one of the main pain points of working with agents, having to explain the project architecture, code style and agreed conventions from scratch every time. Memory is not just a convenience: agents without context tend to repeat the same mistakes and lose the thread of a task when switching between sessions.

Codex as a product has a long history. The first version appeared in 2021 and became the technical foundation of GitHub Copilot, the tool that effectively created the market for AI assistants for developers. OpenAI later relaunched it as a standalone agentic tool. The current update adds a computer-control layer previously implemented in Operator, as the company consolidates scattered technologies into a single product. For developers, this means one agent able to work with code, with the GUI and with memory of the project.

It is not yet known how reliably desktop control works in real, non-trivial conditions. OpenAI acknowledges that some capabilities are at an early stage. The demos look convincing, but the real test is complex scenarios: simultaneous debugging across several windows, testing with visual verification, and working with legacy applications that have no open API.

The race among agentic development tools has reached a new level. Claude Code, Cursor, Devin, Windsurf: all the major players are shaping their answer to one question. Who can carry out a development task from start to finish with minimal human intervention, without breaking what already works? The answer will come not from announcements but from practice in the coming weeks.