I’m having a hard time understanding…
It sounds like this undermines the whole point of having a coding agent… which doesn’t make any sense to me.
I’ll have a look at it later, but I am really confused by this for some reason…
LOL forgive me.
I’ve been using a similar workflow, but I’d frame it less as ChatGPT instead of Codex and more as a separation of responsibilities.
For me, ChatGPT Pro Extended is strongest as the planning, architecture, and review layer. It is where I work through intent, scope, sequencing, risk, acceptance criteria, and whether the plan actually makes product sense. Then Codex is strongest as the execution layer: apply the plan, modify the repo, run checks, produce diffs, and surface evidence.
Codex Plan Mode is useful too, especially for execution-adjacent planning. For example, I recently used it to help turn an architecture document into epics, stories, and tasks. That worked well because the task was bounded and close to implementation. But I still would not treat Codex Plan Mode as automatically better than ChatGPT Pro for higher-level planning.
The biggest lesson for me is that Codex can implement a plan incorrectly or satisfy the surface wording while missing the real intent. A separate ChatGPT/human review pass has caught issues for me that I would not want merged blindly.
So, the winning workflow has been:
ChatGPT Pro / Extended for product and architecture planning.
Codex Plan Mode when tactical decomposition is useful.
Codex for implementation.
ChatGPT/human review before accepting the output.
In other words, I don’t see it as ChatGPT vs Codex. I see it as architect/reviewer plus implementer. The leverage comes from not asking one tool to play every role.
Really? That could be really good in my situation - like others in this thread, I’m doing a lot of copying and pasting, driven by ChatGPT.
I’ll explain briefly how i am using ChatGPT and Codex, but I should preface this by saying I am not a coder (unless you count TMS9900 assembler 40+ years ago), but I do have a technical background, so I have no problem at all following instructions and editing code.
I’m currently modernising a very old vBulletin forum, so it can be moved to a modern server running current versions of PHP and MariaDB, (it’s currently running on PHP5.2 / mySQL 5.6).
With the help of ChatGPT 5.5 (Extended thinking), I built a LAMP server on a spare computer to use as a staging platform. I’m using git locally to keep track of all changes, with detailed logs, and backups of files and the database. I got the site working on PHP 7.4 /MariaDB 10.6 initially, now it’s working on PHP 8.0 and I’m using Playwright to build a comprehensive test harness.
Codex is installed on the staging server and I have used it to do bulk file editing when it has become apparent that the same syntax problems are cropping up repeatedly and there maybe dozens of files that require the same edits.
ChatGPT has been writing all the prompts for Codex, but it is being incredibly cautious in what it allows Codex to do, which is a bit frustrating because there has been a huge amount of copy and pasting back and forth between the staging server via putty, and the ChatGPT app on my PC.
I’ve been wondering could I have just used Codex to do everything? Or is there a way to interface ChatGPT with the codebase on the staging server, so instead of it instructing me to execute commands could ChatGPT do it directly itself?
tbh at this point a coding agent is required only if your code can’t be executed in the ephemeral ChatGPT environment.
ChatGPT can not execute PHP code or spawn a database inside its ephemeral environment so you will have to test on your side. Codex is absolutely capable of refactoring old code on its own.
For this kind of task I probably would’ve used the Hermes Agent - it is far superior to Codex in long-horizon jobs.