Jules: Google’s Codex Killer?

The AI coding agent space is exploding right now. And we have yet another coding agent. This one is called Jules and it's from Google and they are labeling this as an asynchronous coding agent. Now it's very different from something like Cursor or Windserve, which I think are more geared towards wipe coders. But this is probably for professional developers and we have seen quite a few of these in the recent days. There was a codex agent from OpenAI, Microsoft today announced GitHub Copilot coding agent and we have all seen Devin from Cognition AI. Now these are different from Windsurf and Cursor because in this case the agent is assigned a task and the agent is going to go and complete that task asynchronously rather than babysitting your IDE like Cursor or Windserve. And I think it's more closer to the promise of autonomous coding agents that we have. No, I just got access to Joules, so hopefully it's going to get better from this point and I'm going to be testing it in this video. Now. Before that, let's look at some of the features. So it has GitHub integration very similar to Codex and Devin. And in order to complete and assign tasks, it's going to spin up a virtual machine. Now it can write unit tests or run existing unit tests and also show all of its work based on the plan that it creates. So it has this human in the loop component which is really good. Okay, so it seems like you just give it access to your GitHub, select a specific repo and a branch, and then you assign it a task. And after that it's going to come up with a plan. You as a user approve that plan and then it's going to go off and start implementing that. And at the end you get a PR from Jules. Okay, so this looks very similar in functionality to codecs from OpenAI, but let's give it a shot and let's see how good this thing is. Okay, so here's how the interface is going to look like. You need to connect your GitHub account, which I have already done as a free user. You get five daily tasks, which is pretty good compared to something like OpenAI codecs because you need the Pro account in order to even test that out. So I'm going to assign it a real world task. And here's my open source project called Vervey. It's a voice assistant. The idea was to have the ability to bring in multiple different API providers for different things. So. So for example, you can have speech to text system from Deepgram, but you want to use OpenAI LLM. But for the text to speech part you want to use 11lab. Right. So it has support for multiple different models. Now, in this case I'm using the older Gemini API or SDK. So let's see if Jules will be able to help us move this to the newest Gemini Unified SDK. Okay, so I already have that PR for the Gemini integration. However, this specific PR was using the previous version of Gemini SDK. Okay, so I told it this code base is using an older version of Google Generative AI package, upgraded to the latest Unified Google Gen AI SDK and I'm going to provide the documentation. I don't know if it has access to Internet or not. OpenAI's codecs does not have access to Internet. Okay, so first thing that it asked me is whether I want to enable notifications. Let's enable this. It's going to be spinning up a virtual machine and now it cloned the repo. It's reading all the files. So this is pretty neat. Okay, so it says I'm going to upgrade the requirements. Txt file, which is exactly what we want. Then I think it identified that we're using that specific Gemini model in responses underscore generation. So it's going to change the import statement, verify that the rest of Gemini integration is compatible with the new SDK. Okay, that's going to be interesting. One add a test. Optional, but recommended. Create a simple test case to verify the Gemini responses are still generated correctly after the SDK upgrade. This might involve mocking the API call if an actual API key is not available in the test environment. Okay, and this was one of the things I was thinking, whether it will be able to get that Gemini API key or not, and then submit the change with a commit message that describes the changes. Okay, so the plan sounds solid, especially since it was able to identify where that actual piece of code is. Okay, so it says let me know if you have any feedback on the plan. Otherwise hit approve. So I'm going to just approve the plan and we're going to see what it exactly it does. Okay, so the first thing that I see is that it removed the Google Generative AI package from Requirements. Txt. I don't see if it added the newer package yet. Okay, so now it also updated the response generation PY file and we can actually see what exactly it did. So it added this instead of the previous import. Okay, so it's also updated the Response generation file. It removed the previous version of the API and now added the newer version. Now it created this, which I think I'm also looking at here as well. This is where it actually shows you the code. This seems to be correct. And it also created this chart client. Okay, so it's passing on the chart history. Yeah, the code looks to be correct. I do feel like it probably has access to Internet because even Gemini 2.5 Pro, I have seen it fails to come up with correct code for this new SDK. Either they have upgraded the model to because it is using Gemini 2.5 Pro, but the code changes that it made definitely make sense to me. Now it also added some test cases here as well. So this is pretty neat. Now let's see. It says, ready for review. I have upgraded to the latest Gemini SDK. This change updates the project to use the Unified SDK. And here are the changes that I made. Okay, so I said, do you have access to the Internet and can you look up the latest documentation? So it says, yes, I do have access to the Internet and I can look up the latest documentation. I use this capability earlier to view the Gemini API documentation you provided. Is there something specific you would like me to look up? So this is a really neat feature if it actually has Internet Access, because the OpenAI codecs agent model does not have access, which I feel like is a huge limiting factor, especially if the documentation changes. Right. So this is pretty neat. Now you can just publish that to the branch. Let me actually see what exactly happens. And keep in mind, this whole thing took about two minutes. Simple task, but still pretty neat that you can just assign a task to an AI agent. It does it for you. And you can run multiple tasks in parallel as well. Okay, so it created a whole new branch, and if you go to the branch, we can actually see the commits that it made. Here's the file update, here's the unit test that it added, and then the actual changes to the code base. Right. So pretty neat, actually. Now I can definitely see myself using something like this. Okay. It also has a couple of other features. So you can create Codecast, which is basically audio overviews of your code base. Here's how it works. Codecasts are short audio summaries of the recent activities in your repo, and they are audio change logs for your repositories. So if there is a lot of activity happening in your code base, I think this just gives you a quick audio overview of what different PRs are. What are different changes that have been made and so on and so forth. Okay, I still have to test this thing a lot more, but it seems like it is pretty neat. Anyways, do check it out. It's available for free. You can run five tasks and do let me know how this works on more interesting projects. Okay, this was a quick overview of joules does look interesting from my initial testing. In terms of the capabilities. I think we need more thorough testing. So if you get a chance to play around with it, definitely let me know what your experience is. You get five daily tasks. I just tested on two. I am going to test it further, but it's absolutely amazing that we have access to technologies like these for free. Other companies are charging a lot of money. I hope Google doesn't start doing that. But you get a sense of what these technologies will be able to do. And I'm really happy that we are seeing coding agents which are more focused on developers. Okay, at the end of the video I want to share some of my thoughts on the future of software engineering with these coding agents. And I'm not talking about coding systems like Cursor or Windsurf. I feel like the asynchronous coding agents like this or codecs from OpenAI is going to force us in a new paradigm of software engineering and that will require actual understanding of the code base that you are working with. So you need to personally get really good at programming and coding. Start following the best practices of coding and software engineering in general. And that is the only way that these asynchronous coding agent will be able to help. So anybody who actually learns software engineering probably will be able to survive in this new paradigm. Let me know your own thoughts, but this is pretty amazing and it's great to see everybody is releasing these coding edges. Anyways, I hope you found this video useful. Thanks for watching and as always, see you in the next one.