A command center, not a chat window
On February 2nd, OpenAI released a desktop Codex app for macOS, and it's a useful signal of where AI coding tools are heading. It isn't a chat interface bolted onto a code editor. It's built around orchestration — managing multiple coding agents running in parallel, each in its own isolated worktree, with a unified view so you can check status, review diffs, and queue up the next task without babysitting any single one.
That framing matters. OpenAI is explicitly positioning it as a "command center for agents," which tells you where they think the bottleneck has moved. It's no longer "can the model write good code." It's "can a human effectively supervise several agents working at once without losing track of what each one is doing."
Two model updates in a week
The app launch didn't happen in isolation. GPT-5.3-Codex shipped on February 5th as OpenAI's most capable agentic coding model at the time, reported to be roughly 25% faster for typical Codex tasks. A week later came GPT-5.3-Codex-Spark — a lower-latency variant built for real-time interactive coding, notably OpenAI's first production model deployed on Cerebras hardware, running around 15 times faster than earlier Codex versions for interactive use.
Two models in one release cycle, aimed at two different jobs: Spark for the moments you're pairing with the agent in real time, the standard model for longer, more deliberate autonomous runs.
Local, cloud, or somewhere in between
The interesting architectural detail is how the surfaces connect. The CLI runs locally on your machine. The IDE extension lives inside your editor. Both can offload heavier or longer-running jobs to Codex Cloud — OpenAI's hosted sandbox environment — and then apply the results back to your local files once the work is done. You're not choosing between "local" and "cloud" anymore; you're choosing which surface to start from, and the heavy lifting routes to wherever it runs best.
What this means if you're evaluating tools
If you've been thinking about AI coding tools purely in terms of "which model writes better code," that's increasingly the wrong axis to compare on. The more useful questions are about workflow: Does it support running several tasks in parallel? Does it isolate each agent's changes so a bad run doesn't clobber a good one? Can you review a batch of diffs efficiently, or does every task demand full attention the moment it finishes?
The model quality race isn't over. But the harness — the app, the orchestration, the review workflow around the model — is quickly becoming just as important a differentiator.