Codex Expands From Code Generation to Software Work Coordination

Romain Huet describes Codex’s expanding role less as a code generator than as a place to assign, track, and connect work across a software project. Codex is now a dedicated space inside ChatGPT, alongside a separate work agent, and Huet says it has received more than 150 updates in the past two months.

The central mechanism is GPT-5.6 Sol, which Huet calls OpenAI’s new frontier model and says is available to everyone. Its purpose is to work through difficult problems over longer periods. Ultra mode gives Sol a substantially larger reasoning budget, and Huet presents it as an option for unusually demanding projects—while explicitly warning that it will use a person’s token limits faster.

That larger budget is paired with a /goal command and automatic division of work among subagents. In Huet’s example, the goal is to port an entire iPhone application into a production-ready Kotlin Android application while preserving every feature and flow. Codex splits that objective into concurrent technical, UI-asset, and feature audits without being asked to define those workstreams individually. The interface lets the user inspect individual workers or view the team’s activity together.

Huet extends the same model to day-to-day coordination. A user can ask Codex to find five bugs from Linear, create a separate task and isolated worktree for each, and pin the most critical ones. The resulting tasks carry priority and linked Linear context; Huet says they can also reference one another. The coordinating thread becomes a place where otherwise independent tasks remain organized and connected.

Codex can also import work that began elsewhere in ChatGPT. A user can @mention a prior ChatGPT or Deep Research conversation and bring its context into Codex when ready to build, rather than copying research into a new implementation prompt. A research conversation about student-pilot needs, for example, is referenced directly in a request to start building.

Romain Huet places particular emphasis on GPT-5.6 Sol’s computer and browser use. An “appshot,” captured on Mac by pressing Command-Command, is presented as more than an image: Codex receives both the visible screen and application context. Huet says that lets it operate the simulator itself, rather than merely inspect a screenshot.

For an iPhone flight application, the instruction is to navigate every screen and capture App Store screenshots in English and French. Codex clicks through the simulator in the background while the user can continue using the computer. A task such as collecting localized product screenshots can therefore be delegated to the system’s interaction with the running app.

The browser provides a related route for making changes. A user can open a web application inside Codex, click the precise interface element at issue, and annotate it with a requested revision. In the paper-plane game, the “FLIGHT ENDED” state is annotated with the instruction “More playful and animated”; the revised state reads “OOPS! PAPER JAM!”

That visual instruction sits alongside a conventional code path. Developers can inspect a diff and edit a specific line inline themselves. If they know the desired behavior but not its implementation location, they can describe the change instead—for example, asking Codex to tilt the plane more in steep turns. Huet’s point is that the user can work visually when the issue is apparent in the interface, or intervene directly in code when the exact change is known.

Browser work now also supports applications that require login, including passkeys. Huet presents that as extending browser-based work beyond public previews to authenticated applications.

Romain Huet presents Sites as the route from a finished application to publication. He says Sites was introduced for Enterprise teams last month to create and share internal applications; it is now available to everyone for web applications built with Codex.

The request can be as simple as “Publish this to Sites.” Huet says the service includes hosting, authentication, a persistent database, and file storage. The paper-plane game is deployed at an openai.chatgpt.site address, with a standalone landing page and playable application.

In Huet’s framing, a Codex-built web application can move from an idea to a deployed full-stack application without the user having to think much about infrastructure or hosting.

Romain Huet presents the mobile changes as a way to manage delegated work away from the desk. From a ChatGPT conversation on a phone, users can create, search, open, and manage Codex tasks. Mobile review supports filters for unstaged changes, staged changes, a branch, and branch comparisons. Users can also connect directly to an SSH host, such as a devbox, and start new tasks remotely.

On desktop, Codex now brings pull requests into the task context. Huet says the system knows which branch each task is working on and automatically surfaces an open PR for that branch in the task’s summary pane. A developer can follow the work, address issues, and merge from within Codex. If a check fails or a reviewer comments, they can ask Codex to fix the problem with the relevant branch and PR context already present.

The pull-requests tab is intended to support the rest of the review workflow: exploring code changes, leaving inline comments, completing a review, marking one’s own PR ready, and merging it without switching to GitHub’s website.