Cursor 2.0: Composer and new UX in 12 Minutes

--- type: transcript date: 2025-10-30 youtube_id: GS0mtpDiX08 --- # Transcript: Cursor 2.0: Composer and new UX in 12 Minutes a ton of big updates out of cursor. Today they just launched their very first agent model which they're calling composer which is a very fast model that is just shy of the frontier models of GPT5 as well as set 4.5. If you've been following some of the cursor updates over the past several weeks they did have a model called Cheetah and this was actually an early iteration of this composer model. Now in terms of the UI the biggest update here is actually a dedicated view for this new agent modes. As you'll see in the top lefthand corner, we have agents and then IDE. And what you can do is you can specify to toggle back and forth depending on what you're looking to do. Now, one of the other cool things with this is there is also a built-in browser. Here's a demonstration of that where what you're going to be able to do is if you're working on say a web app, you can go and you can highlight different elements and grab the context of those DOM elements and then you'll be able to see all of the different key aspects of it. things like the computed styles and if you want to say in this case put the buttons on the bottom and have the tokens on the top row you can easily make those targeted type of edits. So now before I dive into a demonstration I want to quickly go over the blog post. Within the blog post we see that composer is a new agent model that's designed specifically for software engineering intelligence as well as speed. Now, the way that this model was trained is during training, they gave Composer a set of production search and editing tools and tasked it with efficiently solving a diverse range of difficult problems. The one thing that really stands out with the model is it's just shy of the Frontier models that are out there, GPD5 as well as Sonnet 4.5, but the speed is considerably faster than those other models. Now, in terms of some of the specifics of the architecture, it is a mixture of experts models. This is in part why the model is as fast as it is and it does also support a long context window of 200,000 tokens. Another interesting part that we can see as a part of the training process is through reinforcement learning. We can see as the model goes through more iterations of trying out different scenarios and we can see that over time with more compute that composer got closer and closer to these frontier models. And last up the other interesting thing that they shared is through reinforcement learning the model actually learned parallel tool calling as well as using search and file read tools more effectively. this instead of actually going and updating your codebase potentially prematurely, it will actually hopefully read through and understand what it needs to do before it actually takes any action. Now, if you are interested in some of the specifics of the blog post, I'll also link this within the description of the video. So, in order to try this out, just make sure that you have restarted cursor to make sure that you have the latest versions within here. You'll see as soon as you open it up, we have cursor composer 1, we have multi- aent interface, and we have run agents in parallels. Let's click to get started here. I'm going to click this. And the first thing that we're greeted with is this new agent mode. Within here, we have get started with the agent layout. We can click through to next, manage multiple agents, chat takes center stage, parallel agents, full editor when you need it, and then unified review. First up, a new feature is you have voice input here. What I can do is I can select and I can make my instructions. What I'll do within here is I'll command shift space. I want to build out a Nex.js application that reads developers digest. Let's have a neo brutalist theme. I want a blog. I want a header as well as a footer. And I also want some interactive creative elements on the homepage. Now, I'll go ahead and send this in. And what you'll notice right off the bat is this model is incredibly fast. The first thing that it did is it broke out the task of what it needed to do. Set up a next.js project, create a new brutalist theme, create a header, footer, blog page, so on and so forth. Within here, it went through and methodically created everything that we need. A package.json, ATS config, a next config, so on and so forth. We have the Tailwind post CSS files, and then we also have the layout components. Basically, all of the different components completely from scratch were built within here. It didn't decide to go ahead and create the template of Nex.js. It instead decided to go ahead and actually create everything entirely from scratch. Now, within this chat panel, the one thing that you can do is you can open and close these, but you can't actually edit the files. If I go and try and edit these first, if I go and accept these, though, now we'll have these within our project. Now that I have that, if I go over to editor, we see all that it created here. I'm going to go ahead and open up my terminal. I'm going to go and install all the packages and then once that's done, we'll go and we'll start our development server. Okay, so here is the first iteration of what it built for us. So we have the homepage here. Obviously there are some components that we can't read. So we have white-on-white text, but in terms of the overall neo brutalist look and feel, we do have that sort of look and feel within here. If I click over to the blog post here, we have it created six sample blog pages. If I click through to one of these, I can see that we do also have those pages. Now, if I click through to these blog pages, I do see that these were all functioning. And this all happened in a number of seconds. Now, in terms of how it actually looks, I do actually have some critiques. We'll circle back and we'll give some further instructions. Now, what I'm going to say is I want to make sure all of the text is accessible. A lot of the colors don't actually work well or pass a contrast ratio that is appropriate. So, let's make sure all of the different text as well as backgrounds are appropriate for everything on the site. Okay. So, I'm going to go ahead and send this in. It's planning the next moves. It's reading through the different files and it's creating the next steps. Within here, we see that it's going to update the colors to meet the accessibility standards. And through here, we see all of the different edits. It's just flying through. Now, in terms of the actual UI, it's very impressive how fast it is. But in terms of the actual speed, it's actually moving so fast that I can't actually really see that well all of the different edits that it's making. That is a testament of its speed. But in terms of the UX of just having these tokens stream out really quickly, while it is very impressive, I'm not sure that it necessarily is the best interface for actually having all of these changes. And this is definitely a little bit of a nitpick, but actually streaming out all the different tokens. While it is very impressive, I'm not actually sure that it is the best UX. Okay, so I'll go ahead and apply these changes and I'll restart our server here. We do have a little bit of an error. We have event handlers cannot be passed to the client side component. In terms of React, this is one thing that you typically don't often want to do. You don't want to actually bind different events. You don't want to actually touch the DOM, so to speak. You want to manage things within state. What I'm going to do is I'm just going to take this error and we're going to just pass it in. I'm going to try not to actually intervene if I don't have to. And I just want to see the capabilities of the model. Like is this going to be a model that I'm going to use reliably or is it going to be something that I use just for certain types of things. So I passed in the error and it immediately identified on how to fix this. It's going to replace the JavaScript event handlers with CSS hover states to keep it a server component. Now if I take a look at our application, here we go. We have our neo brutalist theme. We have these fun hover effects. I don't know what's really going on with our mouse here. We have different animation code. Chip explore. I'm not exactly sure. It's a fun little effect that we have there. In terms of some of the other things here, if I just take a look here, it did basically what I had asked for. Like it looks much more accessible. I can read all the text here. I can see that we have the blog post. Now, the one thing within the blog post, I see that these aren't actually rendering markdown. And I'm going to say I want all of the blog post to be about Typescript. Let's have some code snippets. And specifically, I want to have all of the coding blocks render the markdown. Right now, there is no markdown renderer. Make sure to install that. And now, specifically with this request, one of the things that I asked for is to install the package. So, hopefully it finds the relevant package to make sure that we actually can render the markdown within all of those different blog posts. Within here, we see it's writing an awful lot of code. And this model is incredibly fast. I used a ton of very fast models on Grock or Cerebras. The one thing with those models though is generally speaking, a lot of those models aren't actually models that I use day-to-day for a lot of my main coding tasks. It went by incredibly fast, but it did actually go through and run the command to install this React syntax highlighter. I'll go ahead and I'll click to keep this. And within here, if I go back to our application, here we go. We have these beautiful code blocks. So, right off in terms of the first generation, I didn't love that. You saw it didn't look great. Like the initial sensibilities weren't exactly there. This could just be a prompting thing. If I gave a little bit more instruction, maybe explicitly said have accessible colors or what have you. It could have probably generated this right off the bat. In terms of some of the liberties that it went and took, like it does have some interesting decisions that it made. Like I I can't get over this little hover effect on the cursor here. But in terms of actually how it looks and feels, it is a very fast model. It's definitely going to be a pretty fun model to play around with. And I think the speed of it is definitely something that is quite interesting. while it isn't at the frontier, they're very clear on that. They're not trying to skirt through that. And while it isn't at the frontier, actually having these faster feedback loops to sort of stay in the flow. And now, while the design did come up with some errors and the design wasn't on point, you saw with just very simple sort of nudging in the right direction, it was able to generate it was able to improve things quite substantially here. So, it is pretty interesting in terms of the feedback loop of the speed because you'll be able to iterate very quickly with a model like this. So, that's just a little bit of a demonstration of the latest model. Now, additionally within here, another great feature is they now have a browser built in. What I can do within here is I can have a browser tab or I can connect it to Google Chrome. I'll go and I'll open up a browser tab. What I can do within here is I can actually pass in the local host. The really cool thing with this is if I click this select element button, I can select the header and I can say I don't want the double text on these buttons. Remove the duplicate text shadow. You can hover over the DOM element and you'll see all of the different attributes within here. You can see the style, the class, you can see all the computed styles, the position, it has the inner text. So all of the component pieces, it has that within the context. So within here, I'll say I want to remove the duplicative text effect. I'll go and I'll send that in. And the nice thing with this is I'm passing in the exact context and everything is within the same environment. But is really nice being able to select the DOM element of exactly the piece of context that you want to pass into the model. You can open up the console within the web browser and I can say what are the console logs. I'm actually not sure if cursor does have the context of the console. That would definitely be very helpful. If it does have access to this, if anyone knows, please leave a description within the comments. I'd definitely be interested. Now, the one piece of feedback that I have for this is as you see on here is it defaults to the mobile view. And I do have to expand it quite a ways to sort of get a desktop or tablet type of view. It would be nice to actually be able to set like a zoom percentage as well. Maybe if I was at half screen here and I have the chat panel, being able to actually see what it would look like even though it's smaller at full screen would be something that is helpful instead of it just sort of snapping by default with the chat panel a bit bigger to a mobile type of view. But overall, this is a very helpful feature, especially for web developers. Now, just a couple last things. If I toggle over to the editor panel, we can also see the website within here. So, if you want it front and center, it will act within a tab within cursor here. And additionally, if you want to continue on the conversations that you've had within the agent, you'll be able to access them within the right hand side here. It does have a little bit more flexibility in terms of how you actually want to interact with this. And the nice thing with this is you do have the option to also spin off parallel agents as well. That's pretty much it for this video. Kudos to the team at Cursor for getting out their first agent model as well as some of these nice UX updates. This is definitely a nice improvement within the editor and to have a new model that is actually from cursor is definitely something very interesting to see how it will evolve over time. I think obviously it's not quite at the capability of a model like Sonnet 4.5. They don't shy away from that, but over time I wouldn't be surprised if they do ultimately have one of the best models, if not the best coding models in the world just given what they do. Let me know your thoughts. But otherwise, that's it for this video. If you found this video useful, please like, comment, share, and subscribe. Otherwise, until the next