OpenAI recently introduced a software engineering AI agent that can handle multiple development tasks in parallel, making it a powerful tool for boosting productivity.

You can use it to fix bugs, review code, do refactors, and fix pieces of code in response to user feedback.

The real strength of this tool is about offloading or delegating repetitive tasks, so you can focus your time and energy on more important and thoughtful parts of development.

You can check more information in the link below:

https://openai.com/index/introducing-codex/

Since it became available to ChatGPT Plus users last June 3*, we decided to try it out for Android development and share our first impressions.

Prerequisites

To be able to use codex, you need the following:

• A ChatGPT Plus subscription (to access the AI coding agent)
• A GitHub account
• An existing Android project where you can experiment with suggestions or code generation.

Setting Up

To access Codex, you just need an active ChatGPT Plus subscription. Go to chatgpt.com and look for Codex in the sidebar panel.

Setting it up is pretty straightforward. You can follow the step-by-step instructions provided by Codex, or check out the official documentation for more details.

One important requirement: you’ll need a GitHub account and a repository that Codex can access to interact with your code.

Creating and Running a Task

Once Codex is set up, the typical workflow starts by creating a task. You can do this in two ways: using Ask mode or Code mode.

Ask mode is helpful when you’re trying to understand unfamiliar code or explore how something works. It works like a regular conversation with ChatGPT, guiding you through the codebase.

Code mode is for giving direct instructions. For example, you might ask Codex to write unit tests for a specific file or refactor a function.

After Codex completes the task, it shows a diff view so you can review the changes. It also runs checks using the cloud environment that was configured during setup. This allows it to validate the changes automatically.

If you’re happy with the result, you can create a pull request to your GitHub repository right from the Codex interface.

First Impressions

1. Handling Multiple Tasks in Parallel

Codex running multiple tasks in Parallel

One of the first things I noticed was how convenient it is to prompt multiple tasks and simply wait for the results.

For example, I asked Codex to generate a README.md file for each module and to create specific unit tests per file.

I intentionally broke these into smaller tasks to avoid having one large diff. This approach worked well and made the process more manageable.

Normally, doing these kinds of tasks would take a lot of mental energy. Delegating them to Codex and having them done in parallel felt like a big productivity boost.

2. Environment Limitations for Android

Codex works well for small changes, but I ran into limitations when trying to run or test Android-specific code.

I attempted to set up a custom environment script in the settings, but I wasn’t able to get it working properly for Android at this time.

Codex Environment Settings

This wasn’t a major issue for me since I already have a GitHub Action in place to confirm the changes. It builds the project, runs tests, and applies formatting using Spotless.

Still, it would be great if Codex eventually supports Android environments more smoothly in the future.

3. Commit Base and Rebase Concerns

Another limitation I noticed is that when you start a task, Codex locks in the latest commit as the base. If your branch receives new commits afterward, there’s no way to rebase the task from within Codex.

I tried asking it to rebase, but that didn’t work.

The only option was to either create a new task or rebase locally after the pull request was created.

It’s a bit inconvenient, but considering how much time Codex already saves, manually rebasing isn’t a big problem for now.

4. Incomplete Code and CI Checks

There were a few times when the generated code was incomplete or slightly off.

No package name when creating a file. I wasn’t able to catch the problem.

This is where having a CI setup is incredibly helpful.

My GitHub Action checks whether the project builds, passes tests, and is properly formatted.

This allows me to catch and fix issues right after Codex opens a pull request, making the workflow more dependable even when the output isn’t perfect.

Github Action in Action

5. Fixing Code After Pull Request

You can still update code through Codex even after the pull request is created, which is nice.

However, the same rebase limitation applies. If the base branch changes, you’ll still need to rebase it locally.

In some cases, it’s faster to just fix the code yourself instead of trying to instruct Codex to do it.

6. Overall Productivity

Overall, I felt more productive and efficient using Codex. I was able to complete tasks overnight that would have required hours of focused effort if done manually.

Offloading the routine work let me focus more on reviewing and validating code. Because of that, having a strong CI pipeline becomes even more important.

It acts as your safety net when collaborating with the AI agent and helps ensure everything works as expected.

Created 26 Pull Request using codex

Conclusion

Overall, I think Codex is a great addition to my development workflow. While it’s still early and there are clear limitations, I completely understand that it’s a new tool and has room to grow. Despite that, the experience so far has been impressive, and I’m genuinely excited about how much it can improve my productivity moving forward.

As someone who already uses ChatGPT and GitHub for version control, having Codex integrated into that setup makes everything feel seamless and convenient. There are a few rough edges, like the lack of Android environment support and the inability to rebase tasks directly, but there are workarounds that make it manageable. What matters more is that it’s already saving me time and mental effort on repetitive tasks.

For me, Codex is a game changer. It’s the kind of AI tool that fits naturally into my workflow and helps me focus on higher-level decisions instead of getting stuck. There are so many AI tools popping up today, but it’s important not to get overwhelmed. The key is to try tools that genuinely add value, and for me, Codex does exactly that.

If you’re trying Codex for the first time, I recommend starting with small and clear tasks like:

• Creating a README.md file per module
• Generating JavaDoc for each class or file
• Writing unit tests per file or feature
• Creating module or feature templates

It also helps to structure your project clearly, so Codex can better understand it and isolate each task properly.

You can check out my sample repo where I used Codex:

This article was previously published on proandroiddev.com.