Google Just Dropped Another Coding Agent: Jules 2.0

Alright, so Google finally updated and  launched their asynchronous AI coding agent,   Jules. And you can see, as of four days ago, it's  officially out of beta and launching publicly.    It's now powered by Gemini 2. 5 Pro, which is  really good at coding if you've used it before in   places like Cursor. And with Google AI Pro, which  is their all-in-one subscription AI plan, you get   5x higher limits. And with Google AI Ultra, you  get 20x higher limits. So you can check it out on   this page. We'll be trying it out on a real-world  production codebase. It can integrate with GitHub,   it can see your repos, branch changes, and help  you create a PR. It uses a virtual machine,   like it clones your code, and it verifies  that it works. It has a test suite as well,   as it can run existing tests or create new ones.   And it also shows its work, its plan, and so   forth. And basically, since it's asynchronous, it  runs in the background, kind of like how Cursor   background agents do. Anyways, I'll be trying  it out on a real-world production codebase,   which is my app, Tensor AI. And this basically  helps you stay up-to-date with the latest AI news.    You can download it for Android or iOS using the  link in the description down below. And basically,   you can see right now, like, this is all the AI  news that's happened today. You can listen to a   summary and so forth. And I recently translated  the app into Japanese because I was quite a bit   of a demand from Japan. So what I want to do with  Jules is I want to translate the landing page for   the app into Japanese. And then I'm going to have  it do some more complicated tasks, like adding RSS   feeds and so forth. So basically, starting  out, we're going to go to Jules over here,

press Try Jules. And then I'm going to link it to  my GitHub. So I'm going to connect it to my GitHub   account. And now I selected the relevant repo. So  I'm going to start out with a simpler task. Hey,

basically, can you install Next I18 Next for the  Next. js application? And update the landing page   to be in Japanese as well. Make a file with  the translation keys in English and Japanese.    Auto-detect the language that the user is using.   And show the relevant landing page. And then I can   either press Create Plan over here or press Start  Interactive Plan. So I can deeply understand the   goals before actually generating a plan. So I'm  going to just press Create Plan because this is   a pretty simple task. It should be fine here.   And whilst we're waiting for it, I'm kind of

wondering because Google also released a Gemini  CLI. How this compares to Gemini CLI. Of course,   this runs asynchronously in the background.   And it's using the same model as well. But I   wonder if the system prompt instructions or just  anything else is slightly different from this. Or   even if the teams behind both of these products  are working together in some way. It would be   quite nice if in Gemini CLI, you were able to  fire off this into the background in Jules,   for example. For simpler tasks. But anyway, it  seems the plan is now done. And it auto-approves

if you don't approve it manually over here. Let  me just quickly read through this. And yeah,   this is basically what I would have done as  well. So I'm going to press Approve Plan over

here. But I guess one big difference between  Jules and Gemini CLI. Is that with Gemini CLI,   you have to pay for the Gemini CLI separately as a  separate subscription. So if you go to your Google   Cloud Console, then you'll choose it from over  here. And then I think the cost per license is   $22. 80 per month over here. But of course, Gemini  CLI does have a generous free tier as well. But   the thing with Jules is that if you have the  Google AI or Google AI Pro subscription over here,   which I currently have, or the Google AI Ultra  subscription, then you can use Jules with   higher rate limits as well. But the nice thing  is that you can use both of them for free and   you can try out as well.   And the nice thing is that whilst it's running,   you can see the changes that it's making. So  in the package. json file, it installed these   two dependencies. It also made this brand new  file over here. And yeah, it's quite nice that   you can monitor the progress like this. All right,  you can see it's running into an error where it's

trying to run npm run dev. But since it doesn't  have a. env file, it can't run it. So I told it   to suggest dummy environment variables. And for  some reason, it can't do that. So I'm just going   to say finalize it. You're done. Because it seems  to have generated all the files that are required   for this. But it's slightly odd because there's  a system thing that prevents it from being able   to verify its work because it can't create a . env file. So maybe as part of your code base,   you need a. env. example file when you're  publishing it to GitHub. But yeah, that's   the kind of a thing that you just need to do, it  seems. And now you can see that once it's ready,   you can actually press a button over here. See all  the proposed changes that it has in the same file   structure that your project is in. And then you  can press publish branch. So it made a brand new   branch. And then you can also press publish branch  and generate PR. So I'm going to press this,   publish branch, generate PR. And then if I go  to my GitHub, I should see the PR over here from   Google Labs Jules. And then I can press merge pull  request. And that should then update the landing   page on production after it finishes deploying on  Vercel. Anyways, it seems to have worked because I   merged the PR request into the main branch, had  it deployed in Vercel. And then I loaded it up   on my phone, which is set in Japanese. And you  can see it works completely fine. And meanwhile,   my computer is set in English. So it  works fine in English as well over here,   the landing page. So that was a very simple task.   We're going to give it something more complicated.

And basically in the admin panel of the website,  what I do have is a Twitter list. So these are all   the AI accounts that it's monitoring. And whenever  there's new news on any of these Twitter accounts,   it automatically creates a new post on the  application for these Twitter accounts. So   actually what's happening behind the scenes of the  application is that all of these posts came from   one of these Twitter accounts over here. Now, this  is valuable because like you get the latest news   from the Twitter accounts. But I also wanted to  handle RSS feeds for particular websites. So for   example, the official Google blog have their own  RSS feed over here. And I want like a list of me   being able to add RSS feeds to the application and  for it to like then scrape RSS feeds and find any   new information and then create a post out of it.   So that's slightly more complicated. So we'll see   how it performs because it requires a database  migration. It requires some UI changes. It   requires a somewhat big rewrite of some background  functions. But yeah, let's see how that performs.    And this time I'm going to say start interactive  plan. Press interactive plan over here. And now   I think it's going to look carefully through  the code base and find anything that it needs   clarifying questions on and then use that. And  now you can see after looking for the code base,   it asked me some clarifying questions. So I'm  just going to answer these. All right. So I   just came back from lunch and I just checked over  the code and it looks pretty good. It did make a   small error where it didn't edit the declarative  schema that I did say. But that is a rule in my   cursor rules and also my claude. md file, which I  use with Claude Code. And it would be quite nice   if Jules also followed the rules in those folders  or files as well. Because nowadays it's getting   pretty annoying with having a separate like rules  folder or agent. md folder or claude. md folder   for every single agent that there is coding agent  wise. Anyway, we're going to test out the solution   because whilst it does look good on the surface,  there may be a small mistake that I just can't   see. So I'll press publish PR over here. And then  I will switch to the RSS monitoring branch over   here. So you can see over here my claude. md file.   I wish it followed this, this flow for being able   to generate migrations by first editing the schema  and then actually making the migration. But it   seems it would have to install the Supabase CLI or  something like that, which I'm not sure it is able   to do. And when running the function, it does seem  to be working. So it fetches all the links from   the RSS feed and then it starts scraping the links  for the content. And if I go to my link fetch,   it is actually fetching the links and storing  the information in the database. And then that   will be aggregated and put into like AI news. So  overall, I think it did a pretty good job when   reading through the code over here. I was quite  surprised because it also added comments such as   it added over here. This step is no longer  required as recent links contains all necessary   data rather than removing it completely. And  I think that's quite nice because when you're   checking through the code, like sometimes you  forget the logic and it explains it for you in   this case, especially when removing things. The  only real issue is it did not use any mycloud. md   files or cursor rules files to figure out how to  make migrations and use the migration workflow.    And it also did miss a closing tag for one of  the divs which I showed earlier. But overall,

it did a pretty good job. And it's nice to see  that it's bundled with the Google AI Pro plan and   Google AI Ultra plan. So if you're already using  one of these, you don't need to get another plan,   which is quite nice. I know the plans are really  racking up nowadays because people are paying for   like cursor, cloud code, chat GPT, which includes  Codex as well, Google AI plan as well. And there   are so many plans these days. So it's nice to see  this is kind of bundled with the existing plan.    And also because you're not paying on a token  basis, I think it may be cheap in many cases.    I think this is suited towards a much smaller  task, but it did do quite well on a bigger task   as well. It did have a really good understanding  of the code base and knowing where certain things   are. I found that to be the case with Gemini CLI  as well. I think Gemini 2. 5 Pro, which both of   them are using, just has a good understanding of  the code base overall. I think my personal use   case for this would be I wouldn't make any new  features with this. I think that can be quite   difficult. I don't really like making new features  in the background either. I'm probably going to   use this for like adding on to existing features,  which are already well established or doing quick   bug fixes. And one of the nice things about it is  that it's responsive on mobile. So you can click   on any of the chats on mobile over here. And then  you can also see the like chat itself and also the   code and then create the PR for mobile. So you can  monitor how things are going on the go or you can   just make a new session on the go and then come  back to it later on. So I'm actually going to see   the new Jurassic Park, Jurassic World movie with  a friend. And I might actually be using this in   the cinema during any boring scenes, just setting  off new tasks for simple fixes and also doing bug   fixes on the go. For example, someone sent me an  email saying that the upload failed when it came   to updating their profile picture. So I could just  take this error message, then give it to Jules on   my phone and then come back later on my computer  and it should be fixed. I will be using it more   over the coming days. And if there's anything  worth sharing, then I will make a new video about   it. But if there isn't a whole lot worth sharing,  then I'll just leave a comment down below and you   can check out that comment, which should be pinned  somewhere. Anyways, if you have enjoyed the video,   then do subscribe because it's good feedback for  the algorithm that this video is good. And if you   have tried this out, then do leave a comment  down below sharing your thoughts because I'm   interested in what other people are thinking  and how they're finding this tool as well.