Notes ·
Codex Finally Feels Like It Belongs on Linux
OpenAI's current Codex direction is interesting because Codex is becoming less of a standalone coding tool and more of the same agent available wherever development happens: ChatGPT, the editor and the terminal.
Linux is particularly important here.
Codex already executes locally on Linux with native sandboxing built around seccomp and Landlock rather than treating Linux as something that needs to run through another operating system. On Windows, OpenAI supports either its native sandbox or Linux through WSL.
That feels like the right direction for a development agent. A huge amount of real software development, infrastructure and production work ultimately lives on Linux.
The broader Codex experience is becoming more interesting too.
OpenAI now describes Codex as supporting persistent work across ChatGPT, editors and terminals, while longer-running jobs can be delegated and scheduled for things like issue triage, CI/CD and monitoring. Codex can also be accessed remotely from the ChatGPT mobile app, including reviewing output, changing direction and approving commands while the actual files and development environment remain on the connected machine.
Those are meaningful improvements over a terminal-only agent.
History matters. A CLI transcript technically contains what happened, but that is not the same as having a conversation you can easily reopen and continue.
Mobility matters. Being able to start something at a computer and later check it from a phone changes how long-running agent work fits into the day.
Background work matters. Issue triage, monitoring and repetitive maintenance are exactly the kinds of things an agent should be able to handle without somebody sitting in front of a terminal waiting for it. OpenAI's Codex Automations already supports recurring scheduled work.
I am a regular Codex user, but I am not rushing to install every new version anymore.
I've had enough recent experiences with new models and tools regressing that I would rather let them get battle-tested for a few weeks or months first.
That is especially true on Windows.
When Codex was a standalone application, it worked extremely well for me. Since it was folded into the newer ChatGPT application, the experience on my machine has been noticeably heavier and slower. Right now the new application is using around 1.27 GB of RAM, while the older application, now called ChatGPT Classic, is sitting around 478 MB under similar conditions.
I have a powerful enough machine that this is not preventing me from working, but the difference is obvious.
For now I still use Codex regularly, usually alongside other tools when I want another opinion. What interests me about this release is less another interface and more the direction: one persistent coding agent that can follow the work across the terminal, editor, desktop and eventually wherever I happen to be.
Linux deserves to be a first-class part of that.