OpenAI Codex Desktop Application Abandons Built-in Editor
Every founder Dan Shipper published a lengthy article titled "How Codex Happened," reflecting on the journey of the OpenAI coding agent from a cloud product that cooled off to becoming the foundation of the ChatGPT desktop application.
Codex team leader Thibault Sottiaux (Tibo) decided to develop a native desktop application after improvements in model capabilities, rather than continuing with built-in chat forms like Cursor and other IDEs. He believes that the duration of tasks the model can complete has increased, and users will need to manage multiple agents simultaneously, requiring a richer interface than a pure text terminal.
After designer and engineer Andrew Ambrosino joined, the team referenced chat software layouts: a task list on the left, with each task corresponding to an agent; and a conversation on the right. They did not fork the VS Code editor but chose not to include a code editor, allowing users to modify code only through agents.
Ambrosino stated that once manual editing is allowed, users would expect full editor functionality, which the four-person team could not achieve. Before the release, internal usage was low, with some colleagues complaining that simple replacements required prompting the agent. After Google released Antigravity (combining agents with a full editor), Ambrosino questioned the decision.
After the release of GPT-5.1-Codex-Max in November 2025 and GPT-5.2-Codex in December, the model could code continuously for hours. After December, internal feedback improved, and during Christmas, developers completed a full application using Opus 4.5 and GPT-5.2. The Codex macOS application was released in February 2026, with internal researchers Noam Brown and Brockman switching to use it. Altman called it one of OpenAI's most beloved products internally.
Previously, Codex Cloud was popular internally, but external usage was limited. The desktop application became the basis for the new version of the ChatGPT desktop application in July 2026.
Source: Public Information
ABAB AI Insight
OpenAI has set programming as a company goal by the end of 2024, with multiple teams exploring in parallel. After the release of Codex Cloud in May 2025, external adoption was limited, prompting the team to shift towards local and desktop forms. Tibo pushed for an independent model line, allowing Codex model iterations to be unconstrained by the user scale of ChatGPT.
On the capital path, the team rapidly iterated four product forms on a small scale: cloud, terminal, IDE sidebar, and no-editor desktop application. Resources were focused on model capabilities and interface rather than replicating full editor functionality. After model improvements, internal usage increased, and the final product was integrated into the ChatGPT desktop application.
This is similar to early IDE tools relying on manual editing when model capabilities were insufficient, and the interface form shifted once model capabilities crossed a threshold. We are currently in a phase where agent task durations are expanding from minutes to hours.
Essentially, this is a technological substitution: as the model's continuous execution capabilities improve, the necessity for manual code editing decreases. Mechanically, the interface design constrains user paths, forcing interactions to be completed through agents; when model reliability crosses a threshold, user dependence on traditional editor functionalities decreases, and the product form shifts from embedded tools to an independent command interface.
ABAB News · Cognitive Laws
When capabilities have not reached the threshold, the interface must retain manual exit points.
Constrained paths define products earlier than piling on features.
Internal popularity does not equal external adoption; forms need to be revalidated.