Cursor AI Agents Work Like 10 Developers (Cursor VP Live Demo)

There's a way to use Cursor where you have AI agents working for you. AI agents that do your security, AI agents that do your bugs. If you're anything like me, you've tried Cursor, you've used Cursor, but you're not really getting the most out of it. So, I brought on Lee, who is from the Cursor team, to give us an in-depth tutorial with how to squeeze the most out of Cursor. This is for anyone who wants to use Cursor, get the most out of it, have those AI agents working for you. Enjoy the episode. >> All right, we got Lee on the pod from the cursor team. And Lee, what are we going to learn today? >> Yeah, I want to show some practical tips about how I use cursor, how the cursor team builds cursor with cursor, and hopefully things you can take back and apply to building your next app. And so if someone sticks to the end of this episode, what what what are they going to take away? >> My hope is I've been coding for 15 years and hopefully you can steal some of my tips for how I structure my apps, some of the rules and cursor configuration that I use to help you get more out of AI agents. >> Okay, great. Cuz I think a lot of the people listening, we've, you know, we've tried cursor. It's probably part of our workflow now, but we want to get the most out of it. You know, we want to we want to squeeze that juice. And so my hope today is that you're going to help us look, you know, teach us the tips from the inside of how to squeeze the most amount of juice out of Cursor specifically with these agents. Um, and and just, you know, help us avoid some pitfalls. >> Absolutely. Let's do it. >> All right, let's rock. >> Sweet. Okay, let me share my screen. All right, so the first thing I want to show is using agents. Now, the view I'm showing right now is something we're kind of cooking up and will be released soon, but it's basically exactly the same as pulling up the agent sidebar on the right. The reason I wanted to show this is because it's easy to see all of the different agents that I ran. So, yesterday I'm building a real app working on a new course platform and documentation site for Cursor. And if you look on the left, I fired off about 20 different agents, 20 different conversations over the day. Actually, I was coding at midnight last night. So, there's a few up here at the top that leaked into today. And I want to just talk through some of the common patterns I used here to get the most out of agents in this codebase. So, let me just um maybe I'll I'll pick on a few of these and we'll see there's one that's interesting. Maybe uh this event tracking one. So, I asked it to add event tracking when uh the add to cursor button for an MCP was clicked. And you're going to notice that it goes through, it thinks a little bit, and then it finds that there's no llinter errors. So I've taken some measures ahead of time to make sure that the agent can basically fix its own outputs inside of this codebase, which I can show in a second. I'm using TypeScript. I have a llinter setup. I have formatting set up. I have tests and all of those things. The agent can read its own outputs and then just self-correct and fix itself without me having to do anything. So I just say, "Go do this task." Notice it's it's not really that complex of a prompt. I just said very directly, here's what you need to do. And then we can look and yeah, it did a pretty decent job of what I wanted it to do. Now, you notice I got 20 different conversations. Most of these are what, 30 line changes, 75 line changes, 45 line changes. It's important that you're making new chats or new conversations for each kind of discrete task because every single time you're giving context to the model. You can see here it was about 17% of the context window. This is like kind of like the maximum working memory that the AI agent can keep in its brain. So you don't want to clog that up. If you talk to me for 30 minutes, you're probably going to forget something that I said at the beginning of the 30-minute conversation. So for each new task to get the best quality out of working with the model, you can just start new chats for each thing that you want to do. So when you look through here, like it's not that I'm doing super complex detailed prompts. I'm kind of just asking like, hey, looks like there was some layout shift on this video. Let's fix this script that I have. You know, let's replace this uh prompt here. We want to just remove this custom component. Remove this file. we got this error in the editor and we said um could not find the file declaration for this module and I needed I needed to resolve how that worked and it went and said okay you need to add this type in your file. So really the prompts are not too detailed but what you'll notice that I do is that I explicitly tag in files when I want to pull them into the context. I want to stuff them in the agent's brain. So, when I said when you go to a new item in our kind of left sidebar that has all the different links in our course, our documentation site, uh, it should reset the filter input and reset the sidebar outline to the starting state. And so, I'm just very directly telling it what it should be doing. And it goes through, it thinks, it reads the files, it makes some changes, and then again, it reads the llinter outputs and it says, "Okay, the changes that the agent made are correct. We we've got the right code. We know that it's passed all the checks and balances. >> If you like if you and this is this is actually a broader question like if you've never if you're not a developer like should you even use this? >> Should you even use uh cursor or llinters or >> both? >> Yeah. If you're not a developer, there's kind of a spectrum of I don't want to look at the code, right? or I'm interested in just kind of building an app and I do want to take over the code and actually start to learn what it means to build products, build software. There's a growing number of tools I think that are making it easier for people who never really want to look at the code to build software. And I think that that will exist and grow and people will do that. But from what I've seen, you kind of reach a point where eventually you kind of have to look at the code and and it's going to make your life a lot easier as you start to figure out how the code works. You don't need to be an expert. Um, but you can start to kind of tune those skills and it gives you a superpower that previously was just a little bit harder to do. It's like you peeled back the layers of the abstraction to the code. The code is the source of truth and now you can just do whatever you want. >> Yeah. It's kind of like looking under the hood of the car. >> 100% works, right? >> The the analogy I like to give is in the future there will always be people who love their cars. you know, they go buy different parts and they pull up the engine and they tweak all of the different, you know, they buy parts online and they change everything about their engine setup or the exhaust or whatever. And then there's the millions and millions of people who just drive cars. And like both are are legit and valid and fine. There's going to be a ton of people who build software and can build products. And then there's also going to be people who, you know, arteasally write code by hand every day and don't use AI and that's the thing that they love. And that's also valid. I think both will coexist. Real quick, I need to tell you about something that's helping me sleep better at night as an entrepreneur. I was running multiple companies with money scattered everywhere. Zero spending visibility and making growth decisions pretty blind. I knew there had to be a better way. And my friend told me about Bracks. So, I switched and now I have mission control for every dollar. spend management and visibility, virtual cards for my team, and earning cash from the same day treasury product. I was leaving so much money on the table, but not anymore. I have a financial operating system for our businesses to make decisions so I can spend smarter and move faster. And it's been a total gamecher. I'm loving Brex so much. I reached out to them and asked them to sponsor the pod. Start sleeping better. Your financial OS is waiting. Brex is giving SIP listeners $7,500 in perks at brex.com/sip. Cool. Um, okay. I took some notes on this. What else? What else do we need to learn? >> Yeah, I want to show uh a different screen. So going back to the normal kind of cursor window window that most of you are familiar with. I have a few rules and commands set up that are helpful in my project. So this one I'm writing a lot of content here like docs and educational material. And I've kind of built this mega prompt of everything that I like about writing. And there's some things on voice and tone. There's some things on, you know, being specific. But then I also just built this list of banned words. And I'm I'm sure I've cribbed some of these from different places online or books that I've read, but I've just tried to like delete the like marketing speak out of people's brains. So like missionritical, like performant, uh, seamless, like some of these things are fine and all rules can be broken, right? But this is just a helpful list of me trying to delete some of these overused words or overused phrases. And my most recent addition here is also trying to catalog some of the LLM patterns like the junk that it spits out. Like this is the most recent one I've seen people talking about like it's not just X, it's Y. And like once you see it, you can't unsee it where the LLMs all generate the same style of writing. So, uh, this has been helpful for me as kind of just a quick pass when I'm writing something. It's like, hey, am I am I making sure that I'm not doing anything, you know, completely robot sounding? >> Yeah, this is really smart. I've I should do this because I do a lot of writing. Um, and I do this manually right now. So, the fact that this can just be, you know, somewhat automated is really big. Beyond writing, what are other, you know, common use cases that you can, you know, use rules for? >> Yeah. Yeah. I'm happy to send you this prompt. By the way, another one is custom commands. So, this is a little bit more of a detailed command for things that I like to do when I'm reviewing code. So, if you're if you're newer to coding, some of these things might seem a little overwhelming, but you could kind of dial this up or dial this down. The idea holds true, though, which is that you have the agent, and inside of here now, I can run slashcode review. So, I've got this slash command, and that slash command is just driven by this file, this markdown file. We can make whatever we want. We could do code review, security review, um, vibe check, like we could we could do whatever we want here. And inside of my code review custom command here, it's just little gotchas that have kind of bit me in the past. Like, hey, did you think about what would happen when there's no internet? Did you think about what would happen when the application is loading information? Did you handle that with like a nice loading spinner? Did you did you write good tests? Did you write like smaller number of tests that are really good versus just spamming a bunch of garbage? Did you change the authentication? If you did, you might want to run and do a security review. Uh did you have, you know, places where we might want to speed things up with caching like some of these more advanced um software engineering terms and concepts? And just building these own kind of customized commands and customized prompts for your use case has been very helpful for me as well. Yeah, I'd imagine in the future, you don't have to quote me on this. Um, in the future, you would think that a lot of these like agents, so to speak, would come like stock with cursor. >> Yes. >> Um, because so many projects require >> require this. >> Yeah. This is a good time I can show you a couple other things which is one of these kind of code review bots. We used it so much internally just cursor building cursor for almost a year that we decided to actually turn it into a product and it's called bugbot and it allows you to do essentially what we just talked about. The code review prompt that I had was more of kind of my own preference, not just kind of finding bugs. But we have uh another one called bugbot, which just to give the context of this repo, I've got a docs repo for for cursor. Uh I wanted to add some docs about extensions. So I actually kicked this off with an agent and if I want to make any follow-up changes, I can just tag at cursor on GitHub and it will just go make the changes for me. But then I got this comment from Bugbot and Bugbot from Cursor was like, "Hey, by the way, I think you missed some of the other languages." So we have our docs internationalized to a bunch of different languages and I was like, "You missed you missed a few, bro." So I need to go through and pick some of those things. So this is the built-in code review bot basically, which you can make one uh you can make some that are custom. So with our CLI, which I can show in a second, you can run the CLI in your GitHub actions or whatever your deployment uh platform of choice, your CI platform of choice is. So this one, for example, you could audit your repository for security vulnerabilities and you just dump the CLI into this CI script and you can run the cursor agent. So the agent that I just showed that was in the cursor editor, the guey, you can run it headlessly. You can run it just behind the scenes through the CLI. It's the same the same logic, the same idea, the same power, but now I'm giving this prompt and I'm running it basically in automation and I can just automate away and do whatever I want. So you could automatically update your docs every time somebody on the team merges a PR. You could automatically fix things and push commits back to your branch if the CI fails. So if your test failed or something else fails. And this is just like a few of the ideas. There's so many different things you can do with this where you can have complete control and automation over what you're doing. And another thing I wanted to show is uh here was another example of where I used an agent to kick off a PR and I actually went ahead and included the screenshot down below. I got a report on X that this button didn't work. So I'm in Slack and I just kind of pasted in what they said and then I said at cursor here's the repo. Investigate this bug and fix it. And then it comes back and it says, "Okay, here's the fix." I go out to the PR. I look at the code. I go test it. I'm like, "Yeah, seems good." And it just saved me so much time. I didn't have to go open up the editor. I just kicked it off from Slack on my phone. It was amazing. >> Yeah. Like to first of all, I I actually had no idea this existed. I don't know if this is new or not. Um but this idea like this is this is cool because this is about a lot of people are vibe coding, a lot of people are prototyping but you know prototypes are cool but we you know we want to create people listening to this want to create production ready code and a big part of that is >> um you know software that scale software that's secure um being quick on you know when you do have customers being quick on fixing bugs. So I think >> what's cool about this is it um it gets us closer to there. It gets us there. >> Yeah. And the idea is basically that same cursor agent you can use in the editor, you can use it from Slack, you can use it uh from the web. So if you go to cursor.com/agents, same idea. I can just go here and kick off prompts and they can run in the cloud. So they run in this secure virtual sandbox. In the same way I can do it in the editor when I have local files on my machine. So I'm on the go. I'm like in the car and I can just go to curs.com/ aents and kick things off. Similar to other of the like vibe coding prototyping tools. Similar idea here where I can integrate with all of our kind of production repos and just fix little things like this like typo changes, docs issues, the padding's wrong. I want to change the color. Like little things like that to your point really helps with having great customer service and being able to turn around bug fixes super super fast. I can try to find and pull up um the the tweet from Linear, but Linear has this amazing policy where Oh, here it is. They have this amazing policy where they have a zero bug policy and they use cursor. If you look, this is inside of linear. Let's see if I can make this a little bigger. Oh, maybe not. Well, uh PM reports a bug. The AI categorizes it. The engineer who's triaging just tells cursor, "Hey, investigate this and fix it." It opens up the PR, they fix it, and they accept the change, and they just turn this around just incredibly fast for basically any issue that gets reported. And I think a lot of teams are going to start operating in this way. This isn't going to solve like it's not going to say just build me a billion dollar SAS, make no mistakes, but this is like bugs, small things, you know, little things, the the drudge work of building software. Well, I also think that, you know, if I'm, you know, if I'm a customer and I'm on a I'm on Linear's website and I see that, you know, Linear has a zero bug policy and I'm, you know, comparing Jira to Linear and I see zerobug policy, I might be more likely to buy the software that says that >> 100%. >> So, there's like real value, right? Like there's real like positioning and marketing value when you when you actually say something like that, which is huge, alpha, >> a million%. And the last thing I'll just show quickly is like I mentioned, we have the headless way you can use cursor agent. You can also use this in the CLI. If you're the type of person who prefers like Vim or you prefer Jet Brains or you prefer other editors or maybe you just really like the terminal, I think more advanced engineers love this stuff. So I could say in here like let's add uh end to end test with whatever tool you want. I could just say playright or you could just not even say what tool. uh for testing submitting prompts in the chat. Sure. And it's it's the same idea, right? It's it it looks similar, it feels similar. It it has the ability to work with basically any AI model, call different tools, which are essentially like skills. So you you give the model skills to go read files, write files, search your codebase, look up things on the web. It's like you're training it with all these skills and then it can go and make changes. uh you give it this kind of ambiguous or maybe not ambiguous but this broad task like just go add tests make it good make it make it great and it can go and kind of figure out what you want to do it can run terminal commands it says hey what do you want to do you want to allow kind of installing things I'm just going to put it on yolo mode run all commands great just do whatever you want that's fine this is like full vibe um and it can manage the to-dos it's going to go install the dependencies set up the configuration file so very similar thing here with what I was showing kind of in the the visual editor. >> This is cool. This is very cool. I mean, I personally >> I'm I'm not super advanced. So, personally, this is like a bit a little bit It's cool to see, but a little bit daunting to me personally. >> Yeah, I feel that. >> Um, but like I know that there's, you know, I'll call them real developers like see, you know, who who do feel more comfortable in an environment like that. Yeah, it's I personally do not daily code with a CLI. I I like being able to stop and review the code in the editor, kind of click through the files. That's just the way I've built software forever and it still clicks with my brain, but some people definitely enjoy that way for sure. >> Before we head out, is there anything we're missing around if someone is going to use cursor today, you know, how how can they get the most out of it? I think that for a lot of people they have this tendency to make their conversation this append only they just continue in one chat forever and I don't think they realize I noticed this with a lot of people getting started people who are just kind of graduating from vibe coding into building a real production app and they're like it was great at first but then it got bad and they're trying to figure out like why why did it get bad did it get bad because there's just a lot more code now did they secretly make the model's worse? Is there some conspiracy? And almost always when I work with uh people and and try to debug what went wrong, they're just trying to give the model way too much stuff. So like I mentioned with that context window, if it gets up to 100% cursor can automatically summarize and and keep going, but you you don't really actually want to do that because as you if you get to like 80 or 90%, it's just very likely that there's so much junk in there, so much context that the model can get confused. And there's actually been some early research that has shown uh in long context tasks, especially with models that can have a million tokens of context or just a ton of stuff. You don't really want to be kind of running that as your daily way of coding where it's like ah just give it 500,000 tokens of context. It just gets really confused and they they have measured this kind of quality drop off as you get further and further up that uh chain. So I would just be very mindful about how much context you're providing and for new tasks for new features start a new chat you know start a new window and kind of start from scratch again. >> What's the craziest thing you've seen built with cursor? Like in terms of software you know I've seen some tweets people being like oh people using curs you know platforms like cursor they can only build you know little software apps. Is that true or can can you build you know some serious software apps? Yeah, I think a few that have been fun for me. One is is seeing the new generation of people who can just build software now that previously wouldn't have been able to do that and see what they come up with that is just completely it's it's not what I would have ever imagined. So, we did a meet up a couple days ago and a few people brought their parents, their parents brought their kids to the meetups and there was a six-year-old kid coding with cursor which is just mind-blowing to me. I was stacking Legos and like sucking on my thumb or something when I was 6 years old. I don't know. And like six-year-old like building with AI is just amazing. Um, so that's one thing. The other thing that I've seen that's amazing when the marginal cost of writing software is zero, when coding is easy, when coding you can just generate a bunch of code, there are certain types of things you can build that previously you would just never do that. A great example is very very detailed and thorough debugging tools. So you run into a problem in your app and sure you can like add some logging. You can try to step through all of the editor things or you can ask the AI to write a complete visual interface for you to debug and figure out what the heck is going wrong. Now, this doesn't make sense for every single issue that you run into, but I've now seen a couple instances where people built these like custom heads up displays where they could step through what was going on the application or they built this like side byside AB test thing where they were migrating from app A to app B and they could like review each change along the way. You would never build custom software for that in the P. Like it it's throwaway. You would use it once and then be done with it. But when writing code isn't the hard part, it's like, well, heck yeah, I want a visual guey to watch this migration. That sounds awesome. >> Totally. And I also think I mean this whole era of they call it personal software, right? >> Yeah. >> So, you know, this idea that yeah, you might just want to create this disposable piece of software for this one use case and and it makes your life easier so you're doing it. Um, there's also this idea that there's there's probably a really big business to be created in creating personal software for companies too, right? Like a lot of people listen to this show, startup ideas podcast, so a lot of people here have ideas and >> you know I I think that if you're trying to, you know, one of the the simplest businesses to start is go and just use cursor and build personal software for for some vertical. You know, is it healthcare companies? Is it fintech companies? Is it, you know, boring businesses? I think that there's a huge opportunity there. >> Yeah. Quick plug. I actually wrote a blog post about personal software. Robb.com/personal software that talks a little bit about this. It's super inspiring and exciting for me. Like this whole era of people being able to just build things that they want. I've seen so many people build amazing things in this way. And you mentioned that uh you were talking with Logan about Nano Banana from Google. And I've seen people talking on X recently about the capabilities of that model are so good you can just go into cursor and build a vertical of some AI generation image generation app that is going to be 10 times better than anything else that exists. So removing backgrounds or trying on clothes or architecture like before and after photos like things that people who haven't caught up to the speed of technology, they haven't seen the latest versions of the models are like, "Wow, how the heck did they build this? This must have taken so much work." And in reality, it's like it's just a model call now. And you're making it accessible to hundreds of millions more people basically. >> Have you seen Peter Levelvel's tweets over the last 24 hours about Nana Banana? He's like, >> "Yeah, >> his his mind is blown of how good it is, right?" So, and he, you know, can you actually pull up his his axe account? >> Yeah, I'm pretty sure I saw what you're talking about. He was he was dropping some uh amazing knowledge for people just like the thing I love that he does is I think he inspires a lot of people to just think bigger and really try to say you could also be a entrepreneur. You could also just build these amazing things. >> Yeah. I think he just he he saw what you know obviously he's been in the space for a while cuz he's he's been building photo AI. Um so he's been playing with the models and I think he said it's you know Oh my god. >> Show us what's inside Area 51. That's another use case. >> No, it's true. >> Yeah, he had one about um >> Oh, this is a good one. If you're not building a mini startup with Nano Bonano today and launching it tomorrow, you're missing the opportunity of a lifetime. This image model just made hundreds of new startups and apps possible. Your only limit is creativity and how fast you can ship. So true. Like that's banger advice, >> right? So I think like your point is with cursor uh you just call the model like go and I mean you have to come up with a good idea, right? Yeah. Um and you have to have some distribution or understand how to get distribution. But there's this like window right now to create software with this model. >> Yeah. Like imagine this in an app form. It's like a you can drag the map around on the web and then you can like turn it into exactly this fancy um image that you want to see. And then maybe on this image there's different button filters. You can apply different styles. Like maybe there's a a slider that slides it throughout history. So you you do the arrow pointing at San Francisco like downtown on the Embaradero and then you can slide it from 1930 all the way to 2020. And you know back in the day on the Embaradero there was the like over before I think it was the earthquake or something there used to be a like above ground train I think or and that was all along the Embaradero. So you would see that and then you would see it go away and the models are going to know about that like that would be that would be really cool. >> The other thing to think about is you know a lot of people talk about personal software for these utilities but I actually think personal software as distribution software is something that not a lot of people have talked about. So, how do you create a piece of software that, you know, goes viral on X or or gets a lot of traction that uh builds you the distribution that you can basically use it almost as a lead magnet to get and then sell your actual piece of software? Yeah, I actually I'll uh I'll show you something that I I don't know exactly how to put it into a blog post yet, but what you've described I've been sitting on for a while and I've been thinking about and I I've been calling it a distribution engineer. I just keep Apple notes because you know why not? And indie hackers know this, right? They build in public. It's not necessarily about getting followers. It's about getting distribution. And it's the same when you look at basically any other part of building software. If you build docs, well, you want people to read them, right? So, they need to be distributed. If you're building a community, you want it to be more than one person. So, you have to figure out distribution. Distribution is all you need. It's not really about necessarily just the product, but you want the product to actually get used. And there there's something here because in the past, I've talked about how front end and backend went to more product engineers. You were caring a lot more about the product. And now that product is becoming easier and easier to build, there is something here to this distribution angle where okay, I've got a product and I'm I'm an entrepreneur. I'm trying to get it out into the world. Figuring out distribution is huge. It's huge. >> Lee, always a pleasure having you on. You know, you just you're the goat, man. I I'm I'm excited that you're now at Cursor. Um I'm not affiliated with Cursor at all. I just think it's a cool product. Um and uh yeah, man. Let's let's get the comment section going for Lee. I'd love you to come back on. Um do like, you know, people listening, do like this video if you want more of this in your feed. Um and I'll include links to follow Lee on all social platforms. So go ahead and follow him there. Lee, anything you want to leave us with? >> Man, thank you so much. It's been uh a blast chatting and all your recent guests have been fire. So keep up the good work. >> I appreciate it. Uh take care, my man. See you.