OpenAI is methodically assembling a desktop application that would merge ChatGPT, its Codex coding agent, and a web browser. While that unified product isn't launching yet, a substantial Codex update reveals the architectural components being stress-tested for it. Thibault Sottiaux, leading Codex, described the approach as building the super app in public, with this iteration aimed squarely at developers.

The core advancement is a shift to proactive, multi-purpose AI agents capable of operating across a wider system interface. Most notably, Codex agents can now directly interact with other desktop applications. A user can specify a program or let the agent select one. OpenAI contends its technical edge lies in a proprietary layer that manages these interactions without consuming excessive system resources, allowing the agent and developer to work concurrently without significant latency.

This expanded agency is supported by 111 new plugins that bundle skills, app integrations, and Model Context Protocol server connections, effectively giving Codex a broader toolkit and more context sources. A built-in browser includes a commenting system for directing edits to specific webpage elements, and integrated image generation via GPT-4o can produce mockups and assets. Codex can also use screenshots for task verification.

Two memory features are in preview. The first provides task-to-task context retention, which OpenAI states will improve execution speed and quality over time. The second uses accumulated context to suggest proactive actions, like prompting a response to a colleague's document comment at the start of a workday.

The update is rolling out now to desktop users logged into ChatGPT. Computer use is initially available on macOS, with EU and UK access coming later. Those regions will also see a delayed release for the memory functionalities.

Source: Engadget