This weekend... evaluating Codex. It's a mixed bag. I think the workflow is about perfect. I wish worker processes were automatic. They are there, you just don't see a lot of it. Git conflicts are pretty constant with it, unless you approach your task lists very carefully.
And, every time you ask the thing a question, it spins up your entire environment and runs the task runner, which is bad. It shouldn't need to do that. Here, again, I'm finding that the quality of the task list you give it is everything. Long well designed task lists work better than "make me an app that does X" which is fine.
What I don't like here is the way it handles sections on a task list. Three tasks in a section, it does exactly three tasks and reports back. Then, you have to tell it to do the next section, and the one after that, and so on. Running sections concurrently is a mistake, because Codex agents don't share a state. Each one runs in its own vm.
One agent can optionally spin up sub agents as taskrunners within a vm, but it won't happen in a predictable way, and you have to deal with whatever you have to deal with there.
Also, it doesn't self correct, so I'm going back and forth between Codex and windsurf to QA the work.
Thinking, generally, not crazy about the GitHub integration, but I can't think of another way they could do this with a cloud service that makes any sense. I also expected OpenAI to pull that little thing they do, where they limit how much I can do with the tool for no reason out of the blue. So far, they have not done that.
The other thing I'm annoyed with is the system doesn't consistently honor changes in my environmental settings by default. This feels like a bug, and I wish they would fix it.
In all, it's saving me time, helping with planning and QA, and serving, for now, an indispensable function in my workflow.