you can use running Codex separately from the terminal to complete a specific task, where the agent running Codex receives only the final output from the task.
You can imagine that it is actually a small agent used by all code agents. It is simply a very powerful sub-agent that can spawn other sub-agents because it is the Codex itself.
In this article, I will discuss how I use the Codex exec command for maximum productivity, highlighting how I use Claude Code and codex together to have the best code setup.
Why use Codex exec
The first question you might ask is: Why do I have to use Codex exec and not just run Codex myself to complete tasks? The simple answer is that these have very different use cases. You must, of course, run Codex yourself if you want to complete a task with it directly, but the Codex exec command is very useful when using other coding agents.
Usually, I have Claude code that runs the Codex exec command to perform some actions, usually reviewing the work done by Claude Code. I can also use the Codex exec command from another codex agent because the command will start a new thread where the context is not affected by the previous work I did.
This is great if you want to review a job, because you don't want the reviewer to have all the previous context. The whole point of having a reviewer is that they don't have access to all your logs. It simply checks what you've done, compares it to the task you're trying to achieve, and can update if there are bugs or if you didn't complete the task correctly.
How to effectively use Codex exec
Understand what Codex exec is.
There are dozens of use cases for Codex exec. I will discuss specific use cases for which I use Codex exec, but you should also think about your workflow and how you can get the most out of Codex exec.
In simple terms, a Codex exec is just another code agent that you execute with no previous context, and it completes a piece of work given the initial prompt you give it when you run the command.
codex exec "" Codex exec as update agent
For example, you can use it with the following command:
codex exec "Go through this PR review the contents of the PR, look for
any severe bugs, classify them into P1, P2, and P3 level issues.
Also review the original task description and determine if the PR
solves what the task asked for."With this simple command, you've got yourself a powerful update agent. I think this is the most important case to use for Codex, because Codex is an incredibly good code analyzer.
I use Code Claude as my main driver when I do coding. However, I use Codex exec to review the code that Claude produces.
I compared this to the review code and Claude Code, and I don't think it's even remotely comparable. I think Codex is a very powerful analyst. I see this in 2 main ways:
- Codex can see problems that Claude Code can't. This prevents many bugs, and after using Codex as my code reviewer, I have almost completely eliminated bugs caused by new code being added to production, which is an amazing achievement. This is actually a review agent recall.
- Codex, however, is also better at accuracy when it comes to being an update agent. I think that when I use Claude's Code to review the code, it reveals a lot of non-issues that don't really matter or are wrong. I rarely have this happen with Codex, and in the few cases where Codex also makes such a mistake, I have noticed that Claude Code can also make the same mistake.
Overall, Codex is simply an amazing code reviewer, and using it is one of the easiest things you can do to greatly improve the quality of your code and reduce the number of bugs you encounter. When using Codex as a review agent, be sure to correct all Codex feedback before integrating any code into production, and ensure that you re-request a review from Codex after correcting such issues, and continue until Codex approves your PR.
Making this simple change will give you instant improvement.
Codex exec for editing
Another very powerful tip for using Codex exec and getting the most out of it is to use it when planning an implementation.
I think Claude Code has a very powerful and efficient feature, called Ultracode. This basically creates more subagents, uses more tokens, and makes the overall Claude Code work better. I think it works especially well when doing refactoring.
However, editing the Claude Code is not complete, just like the Claude Code review code. And in the same way that I use Codex to review code before putting it together with dev or prod, I will also use Codex to review a program made by Claude Code.
Now you may ask yourself a question: why not just make a program with the Codex itself? And yes, you can definitely do that. I have mixed information about it, and in my experience, the best way right now with the models available is to use the Claude Code to do the programming, but have the Codex review the program, make sure it serves the purpose you are trying to achieve, and that there are no problems, misunderstandings or ambiguities.
In order for Codex to review a plan by Claude Code, I simply describe the task to Claude Code and tell him to create a plan and complete it. Then, when the program is done, the Codex must update the program with some information that we always use and repeat it in the program until the Codex approves the program. You can simply type this into your Claude user-level information or simply tell Claude Code to do it every time you create a new implementation.
Codex issued a second opinion on the topics
The last special case is when I use Codex exec to get a second opinion on a topic. For example, if I determine the architecture decision or find the root of the bug, or similar. I will usually ask Claude Code first and then make sure Claude asks Codex for an opinion.
Now, an alternative to this would be for me to investigate the occurrence of a different Codex and question the Codex immediately the same way I question the Claude Code. This is, of course, very possible and something you can do. The problem is that this takes time and gives you many sessions to hold. It's much better to tell Claude Code to spin the Codex itself, since there's no real difference between Claude spinning the Codex or me spinning the Codex, and the Claude Code just gets the extra power the Codex provides without me spending more time scrolling through Codex sessions, updating the output, copying and pasting the output into Claude Code, and so on.
Simply put:
Having Claude Code spin up Codex sessions using Codex exec is much more efficient than me running Codex sessions myself.
The conclusion
In this article, I discuss how to get the most out of the Codex exec command. The Codex exec command is basically a way to twist the Codex to do a specific piece of work, and it works great for getting a second pair of eyes on a problem. This could be reviewing the code, updating the program for implementation, or getting a second opinion on the topic.
I think the idea I described in this article will be very important in the coming years. We will have several different boundary level agents, coding agents available to us, and we will probably find that some coding agents are better for some jobs and some agents are better for other jobs. For example, now the situation I believe is that Claude Code is better for raw implementation, while Codex is better for reviewing the code and reviewing the generated output.
So, to get the most out of all the models, I use the models where they work best, i.e. Claude doing the main implementation and using Codex to update the code. Integrating LLMs like this is a great way to create powerful coding systems, and I believe this will become even more important in the future.
👋 Touch
👉 My Free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free ebook Vision Language Models
💻 My webinar on Vision Language Models
👉 Find me on social media:
💌 Stack
🐦 X / Twitter
