Cursor for Java Developers

0:00 Uh today we'll be talking about cursor uh the product that everyone loves so much. [laughter] So let's go ahead and get started. Um in the past few years we've been working very hard to make cursor into not only just a IDE but into an ecosystem. So a lot of people know us as just an IDE. So think about tab, think about control K. Uh but we we've evolved a lot. Um in the last year we actually started with CLI tools. Um some background information. I

0:29 talked to a lot of my friends who are also software engineers as well as customers. They tell me that hey I like to use codeex and also cloud code more because they have the CLI. But little did they know that cursor actually has the same thing. We have the CLI and on top of that you'll be able to access all the latest and greatest models at the same time have them debate with each other. So in my opinion it's a better tool to be used and at the same time

0:54 cursor is a harness lab. So each time all the model providers before they provide their latest and greatest model, they actually send it to cursor for cursor to play around and test it out and build a very good hardness for each in specific models. Highly recommend it. Next, cursor also is a front-end design testing tool. You'll be able to use web browsers within the IDE. My favorite feature by far is it's able to access a DOM. Um, and the next thing that we're

1:27 also looking into in the future is hosting cloud agents. So, have you ever kicked off a long running agent task that runs for like 30 minutes and you just got stuck? You can't close your laptop or anything? That's something that will be solved. Right now, you'll be able to go to cursor.com/agents.

1:46 You can kick off any task. Choose a model that you want. Just send it off. An hour or two later, come back. it'll be all completed. And the next thing that we're also very very enthusiastic in sharing is uh automations across all software development life cycle. Uh you'll be able to trigger with linear or trigger with pager duty different things and it'll be able to slack it to you once it's done completing some um task. It's similar to um if people know been following the latest news of like uh

2:23 open claw gas town. It's something that's more contained and much more um efficient so that uh people within the cursor environment is able to use. So here's kind of like a bird's eye view of what cursor is. It has a desktop, you have the agent platform, you have bugbot code review, uh you have cursor CLI, and the most important thing, the reason why that you're in this seminar uh webinar is uh Jetrains ID. Uh last week we launched an integration with ACP with

2:57 Jet Brains. So, we're very excited to start demoing. So, before I get started, let's go go ahead and do a quick quick start. Um, let's see if people in the chat has already done it. Um, go ahead and go to cursor.com/d download. Go ahead and download it. Only takes 10 seconds. Um, and with Intelligj users, uh, if you already have Intelligj, um, installed, let's go ahead and add the ACP. So, um, I'm going I'm going to switch my screen to Intelligj here. Are

3:31 people able to see my Intelligj? Can I see a thumbs up? I see a raised hand. I'm taking that as a yes. >> Sounds good. That's perfect. So, before I start, um, in the chat, can we get, um, let's say if you're a power, uh, powerhouse cursor user, please put a five in the chat. If you're just relatively new, uh, please just put like fresh eyes. First time using cursor, put a one.

4:02 Somewhere in between, put a three, please. So I can tailor this um this presentation here. >> Do you think I can get a quick average? >> Yeah, you're getting quite a few. You get whole range here. Um but a lot of we've had ones to fives a lot around the three range. >> Around the range is sort of our main our main section here. Okay. >> I am I'm going to switch off the chat here afterwards here just because it keeps us focused towards the Q&A as

4:30 well. So uh just for the attendees. >> Yeah. >> Good to know. Good to know. Thank you. Well, um that's that's that's a good understanding. Um today the presentation will will be geared towards um that part. So let's go ahead and start setting up. Um on in my IntelligJ, you can see my file structure over here. These are um the files if I opened anything. On the right hand sidebar, the second icon, you can see AI chat. Let's go ahead and open that. And over here, I already have cursor installed, but for

4:58 those of you that hasn't done it, please follow along and do so. Open this up. Install from ACP register. Let's click on that and simply just type cursor. And you you can see our v 0.1.0 that launched last week is here. Go ahead and click install. You can just chat with it and say hello cursor. And just like any LMS, it' be able to um uh respond back to you and say it's helping me with the Robocode project. Um so quick overview of what we have in this ACP function. You can see that cursor itself has a very cool very

5:40 um exciting function that is semantic parsing. What that means is with your whole code code base it's able to have semantic semantic meaning to it. So if you ask any specific questions of meaning to it, cursor is able to parse out of it not just through GP but also with semantic semantic meaning. Um it's using a vector database to uh speed speed up the development cycle. And next on the list is you'll be able to see your chat history up top. You can see

6:14 all the chats I had with Intelligj. And um down here you can have different modes. You have ask mode, plan mode, and agent mode. Um later on I'll dig deeper into plan mode, but right now let's focus on ask mode here. And on the right hand side you can choose all the models there are um available to you. Um composer 1.5 is by far my favorite model. It's uh it's a new in-house model that was created by cursor um that launched about a month ago. Um it

6:48 handles very well with medium reasoning tasks. I'll I'll go in deeper into I see a lot of hands raised. I'll go in deeper into model selections um in a bit. So before we keep going on, let's let's just try out ask mode because I got into this RoboCode um repository. We don't know what this this repository what does it do. So let's go ahead and say um hey cursor give me an overview of the project and tell me how to start the game.

7:22 And just like that um cursor goes ahead and read through some files. Um it's going to give me an overview overview of the project. It's reading through a bunch of files. It uses GP and what we touch upon which is semantic search. You can see it comes out ultra fast. Um, it gives me a quick understanding of project overview. Robocode is a Java programming game where you can write code for robot tanks. You guys are here are in here for a treat. Uh, that uses

7:51 swing swing gooey. That's very interesting as different models modules. Uh, some requirements that needs to be that that needs to have is JDK 17. That's very good understanding. And how to get started and run your project. Okay, this is good to know. So it seems like in order to to run my project here, I would need to do a build CD into a sandbox and then run this. So that that's that's way too many steps. I I feel like so what I want to do here is

8:25 um change into agent mode. And one thing that's very cool about cursor is within the same chat you'll be able to switch in between modes and models as well. Over here I actually want to create a skill and say something like um please please create a skill that starts the game based on the context that we have with that. Um, I'll show you guys in a second what does the current the skill do. Um, and essentially what the skill does is it creates a MD file um that

9:03 lives undercursor uh skills and then you'll be able to see start game over here. It essentially created a start game sort of like a skill and you can say as a command as well. It's like a atomic action that can be called by not only human beings but also the agent as well. So you can see all the things that it needs to do. It needs to build. It needs to CD into sandbox. It needs to run. Um it it's all included in there. So let's go ahead and try that out. Right. Let's say start game.

9:42 Uh it does it have s? No, it doesn't have s. Start game. It's going to the terminal to run some commands. And it seems like it's spinning up now. And you can see this is our robo code. Um you can see sitting duck and tracker is fighting. can adjust his speed and you can see that. Oh wow. So tracker one. Um let's try something else more. Uh let's let's create a new new game. Um we can put corners. Let's just put the first three, right?

10:30 Corner, crazy, and fire. Let's have some wagers in the chat. See who would win. Let's go ahead and start this battle and see what's up. Okay, corners. It seems like it went to corners. that as expected. Fire is indeed firing. Crazy is just being crazy. You know, it's it's a little bit slow. Let me speed this up a little bit and see who the winner is. Oh, wow. Okay. Okay. So, it seems like fire is first place. That's very interesting. So, are we able to build a

11:05 robot that's able to beat fire? Right. That's that's probably our today's goal. Um in I know there's a lot of um good cursor users over here. Uh people would know that okay the next thing you want to do is open up a new chat because context is quite scrace in LMS. Um it's like similarly when you talk to a human being when you talk to a human being if you talk to them for 30 minutes straight with different topics at all times they would get a little bit confused by the

11:37 30 minute mark. Whereas, if you open up a new chat, it's a completely fresh context window, you might perform a lot better. And research has shown that if you fill up the context window, the model tends to degrade. Doesn't matter how great the model is. So, let's go ahead and open up a new chat. And it's very tempted to just go ahead and do agent mode, but instead, I'm actually going to show you guys plan mode. Let's use plan mode. and plan mode um is

12:07 normally needing some more um a model that has more intelligence that has more reasoning skills. So we all know that GPD 5.4 just dropped recently. Let's go ahead and use that. And let's say something like um investigate the abilities of Firebot and try to create another bot called Water that will beat Firebot in a one-on-one situation.

12:42 Let's go ahead and send this off here. So basically our goal is to beat Firebot because Firebot is the strongest spot that we've seen so far. And what we're trying to do here is use GPD 5.4 in plan mode. So why did we choose GPD 5.4? This is a good time for while the model is working. It's calling sub sub aent while it's doing its work. Um let's go ahead and jump back to our slides actually and It's a model model choice 101. So as large language models are are coming out

13:22 so many of them coming out every two to three months cadence um it's a good way to learn cursor because cursor is a model neutral platform you'll be able to take advantage of um using the frontier models whether that's um opus set uh GP 5.3 GP 5.4 for you'll be able to use that in our platform.

13:47 And how I like to think about model selection is I normally group them into two groups. One is a async and high reasoning group. Um what that means is they're very highly intelligent models. They have strong reasoning efforts and they're tend to be very excellent at long horizon tasks. Um, so that's why when I'm doing the plan mode, I wanted to use a higher level reasoning model.

14:17 That's why I choose GPT GPT 5.4 here. There's a drawback to them, which is they tend to be a little bit more um costly compared to all the other models. Um, over here you can actually see we have two dimensions here that are um classified. One is spend efficiency, the other one is terminal bench score. So um very interestingly terminal bench score as of Q1 2026 the top score is not 5.4 it's actually GPD 5.3 code X and in terms of spend efficiency GPD 5.3 code X

14:55 is actually the highest compared to all of our group ones here GPD 5.4 and also CLA 4.6 six opus. The next group of model selection are fast and synchronous works. These are like the executors. They're like the implementers. They're near frontier intelligence with very fast speed. Um so think about your 4.6 Sonnet. Uh Gemini 3 Flash. That's a good one for documentation. And my personal favorite, Composer 1.5. It's ultra fast and gets things always right within a medium medium level of difficulty. And never

15:36 forget about Haiku. Haiku is also great. Um and you can see the respective terminal bench scores over here. Um you can see composer 1.5 um spend efficiency is great and also the terminal bench score is high as well. Let's slowly look back. Okay, so keep simple. So in plan mode a lot of times they actually ask you to um ask you a lot of questions. Let's let's see uh what is the main goal of water optimize being mainly to be sample.fire at a duel um still keep a reason I I think it's it still needs to have some variety if we

16:15 put in some other agent. So even though it counters fire but at the same time I I still want some more um understanding of it. So, seems like Okay. So, right now, um it's still creating a plan. Let's go ahead and let it let it um keep cooking. Um let's go back to our our slides and while we wait for for it to cook. >> Um and next thing I would like to touch upon is um some Oh, yes. >> You mind? We have a couple questions actually that might be worth calling out before you jump into some of the

16:48 primitives just more around the the IntelligJ connection. >> Oh, yes, of course. that works. Um, one of them is and if we were going to cover this later then feel free to punt it for now but uh a question around MCP servers and whether we can connect MCP servers with the uh with the ACP connection we have here. >> That's a that's a very good question. So um if you're coming from u VS code um or um cursor itself um the MCP is just one click we have our marketplace that's very simple. Um, however, with our ACP

17:22 connection with Intelligj, you actually need to create a MCP.json um at the user root level. So, whatever your user root level is, you want to put your MCP um server there. So, I can I can actually show you guys here um what are some MCPS that are connected.

17:47 And let's go ahead and change this to composer 1.5 because um again this is a relatively fast execution task. We don't need a um high reasoning model for this. So you can see that okay I have Jira MCP that's that's installed uh and sub aent MCP. That's good to know. Um let's see here. So this is where you want to put your MCP.json. Um so there's two things you have to do.

18:16 One is you need to configure your MCP under um this root directory over here.cursor MCP.json. The second gotcha which took me about 30 minutes um the other day to find out is you have to go to in intelligj IDE go to settings type in MCP um in MCP server over here you have to check mark this is a security feature within intelligj that um that that doesn't enable MCP by default so you after you check mark that you'll be able to use or MCPS. Hopefully that answers your question.

18:59 >> And and one quick followup on that too because there's a question around authentication and MCP. I just correct me if we're wrong here, but I think it's just that with authentication, it's the same as using MCPS through cursor. We you can pass in tokens depending on your MCP config or for more like OOTH MCPS then there's a login flow that you can go through for those. Now, correct me if I'm off there, Joseph, but I think that's that's it for that piece. Um

19:24 >> yeah. Yeah, >> that's correct. Um, and sometimes, um, a lot of I I noticed a lot of our devs, they actually use, um, iterm as well. Um, they used in within the CLI, you'll be able Let's Let's actually spin up agents. >> And then you'll be able to do /mcp login and then you'll be able to choose whichever MCP identifier that you had at that time in order to log in with OLA.

19:52 Um but then I think um the most trusty source I still would say the first pass that you want to do is with um API keys under um MCP.json. >> And I'll throw in one quick thing too which is for anyone who's the cursor CLI, the ACP integration here is is powered by that CLI. So a lot of the stuff around how you can configure MCPS and all that kind of thing. Uh a lot of it is is similar or exact the same to the CLI. Um, so that's also something that's worth kind of knowing here is it's it's kind of connecting through

20:23 that same system. >> Yeah, thank thanks for adding that, Sean. Um, so let's actually go back to our uh plan mode over here. Um, so within plan mode, it seems like okay uh we finally we have our build water to counter fire uh robot out. Um you can see that one thing that I love about plan mode it gives you all the detailed things that the model has learned. So cursor learned that this is the things that I learned proposed design files to add implementation these are all good thing and the most important thing is

21:00 to-dos. So you always want to review the to-dos to see if it has too many extra steps. So let me see what's going on here. Okay. Add water property. That looks good. Build and verify. dual check. So I actually don't want it to uh build and verify. I don't want it to dual check. The reason why is because we'll be manually checking in the end. So let me just go ahead and tell tell cursor over here, can you please rem remove the build and verify as well as

21:29 dual check to-dos because we'll be manually verifying. Just code up the waterbot, please. and let's just send that off. Well, it's not scientifically proven that if you use please or thank you is a good thing for the model or not. We are burning some um some some GPUs in the data center but you know it could it could be helpful. So um so in this case um the agent went ahead and start coding. It seems like it coded up water Java um water java. And yeah, it looks pretty

22:13 good. Is it looks pretty good. Let's Let's look at on on our Let's switch tabs into our Java UI, right? Um let's um open Oops. Uh a new battle. Um let's remove all the all the bots over here. Let's put our fire, the one that we want to beat. And let's see here. Um is our is our water. So, it seems like our water is actually not in here. This is interesting. Um, let me let me try and uh let me see here. Let's let's just try to um say, hey, I don't So, one thing that's very cool about cursor is you can

22:58 actually add screenshots. So let's go ahead and um uh provide some some feedback over here to the agent and saying that hey I don't see waterbot in my um in my start game. What's going on? So this is a this is a bug that we see and this is something that happens quite often um in the agentic flow um development um a lot of things are non-deterministic it's probabilistic so a lot of times sometimes things go wrong but what's great about it is with cursor you have

23:41 the resources to provide more context um so it seems like okay water is not in here uh so we need to rebuild and copy water in there. Okay, that's good to know that we actually need to rebuild. Um, so it went ahead and ran this command. Restart Robo game. Okay, so let's see here. Um, let's go ahead and kill this. Um, let's start a new chat.

24:14 And I have a skill that said kill games. I'm just going to reset everything to see um if water shows up afterwards. >> Awesome. Um and if you want, I can toss in a question here while we're having it rebuild. >> Yeah, of course. There was a couple questions that folks had um around kind of differences between this way of using cursor versus the IDE and if there's anything that's like missing from this version of using cursor in Intelligj uh that we have

24:47 inside of the the main ID. >> I see. So the main main ID meaning um the main ID in the cursor itself the desktop app. Yeah. So um one thing that I would like to point out is ACP indeed is um is an extension um cursor app itself is um is a native AI native app that's created. So the flow would be a lot better. So, one thing that I can point out immediately is you can see that within the modes there's agent, plan, and ask. If you're a hardcore cursor user, a five out of five cursor user, you'll realize that hey, where's

25:31 debug mode? And that's one of the things that we're working with um Jet Brains right now to integrate that because we have to hop over some loops. However, the good thing is that the underlying hardness and also the model that's being used within the ACP is exactly the same as the IDE that we use with cursor.

25:56 All right. Um I hope that answers your question. Um so indeed I verified and see that hey our water bottle was here. Um we realized that it's because we didn't rebuild. Thankfully, cursor go ahead and rebuilt it. And now we have a duel between wa water and fire. Let's see. Let's see how this plays out. I'm I'm a little nervous here.

26:21 Interesting. Interesting. It seems like Oh no. Oh no. Wow. Seems like our water bot actually lost. This is This is very sad. So that's that's something that we can go ahead and tweak around and talk to the agent and say, "Hey, here's a here's a result of the score. We actually lost to fire. Can you go ahead and tweak around and make things make things happen?" We can actually um we can do do we have time? We have We're at 32. Um yeah. Um we can we can actually do something. Um there's actually other things that I would like

26:59 to cover. So, next thing that I would like to show you guys is a little bit of the CLI. Um, before I jump into a CLI, um, John, is there other questions that that there there is that I can answer quickly? >> Um, yeah, I think just to make sure, um, maybe I I might just rip through a couple things that have come up a couple times for people.

27:20 >> Um, so there's a couple questions about whether you can use this in Intelligj with uh whether you need to pay for Intelligj uh AI. Uh, they have their AI premium offering there. As long as you have an IntelligJ license to be able to use the IDE, you can use this agent. Uh as you need a cursor account as well too. Uh but you don't need to go for that extra piece there. Um it is helpful if you have it, but it's not something that's required for this integration. Um

27:44 and then the other piece as well too is uh we were talking here about um about Intelligj, but one thing to call out is that this integration is set up to work across other Jetrains idees as well. So if you're a PyCharm user, if you're working with other IDE there, then this is going to be uh also something that you can use. So uh it is available in some other idees there across the Jet Brains lineup and we hope to expand that over time as well too. Um awesome. And

28:11 then I think that's good for now. I think Joseph, what I'll say is there's a few other questions around like rules and skills, but I think we could leave that until after the CLI. Up to you there. >> Yes, sounds good. I I'll keep going with the with the um with the CLI and afterwards when we're waiting um some periods we can jump back. I have some slides around that as well. So um for those of you who are um hardcode uh codecs and also clock code users. Um hey we have a we have a CLI as well. All you

28:40 have to do is here I have three um three terminal tabs open. Um, one thing I like to do a lot is um, within the terminal you have multiple things that you can do. Uh, you can choose different modes by doing shift tab uh, auto run all commands, blam mode, ask mode, debug mode. Here you can have debug mode or just plain agent mode. And here you can also um, one of the questions that we had was create. You can be create rules, you can create skills, you can create

29:12 sub agents or all the other slash commands that you want. Um, next thing that I'd like to showcase you guys is you get to choose different models. So on my left hand side, I actually want to use composer 1.5 in the same in the same directory directory. I can spin up agent and I can say model and let's say um I want GPD 4 5.4 4. Um 5.4. Let's do fast. Uh let's just do let's just do 5.4. Um high is fine, too. Um actually, yeah, that's fine.

29:50 [laughter] And then um in in the lower terminal, I'll be able to say um agent and I can say model and I can have opus 4 point uh opus opus 4.6. So within the same directory, I'm actually able to have three different models that's open. And that's something that's very powerful because let's say today I want to work on three separate tasks. Um, and one thing that I want to showcase that's pretty cool is we can say something like use your best judgment and create the best bot that can win in a one-on-one battle and name

30:33 it cursor because this is a composer 1.5 uh composer 1.5 model. So then I'm going to name it cursor. Um, this is GPD 5.4. So, I'm going to name it GPT. And then this one over here is Opus. So, I'm just going to name it um Opus. So, in this case, you're kicking off three parallel agents that goes ahead and create three different bots for you at the same time. So this is very powerful um in the CLI and you can do this in the IDE as well. You'll just have to create new chats and manage it

31:21 through your history over here as well. Um so that's that's kind of like a soft plug with our CLI and once the agents are done um compiling we'll be able to test it out and see the the battle between cursor opus and GPT. That would be fun. Um so while it's implementing this um let's jump back in um understanding the primitives of rules skills and sub agents. Uh one of the questions that we had so um we've been seeing within cursor we've been seeing

31:56 this year that we've been change software development has been changing from writing code into orchestrating a team of AIS. So when when it's combined, cursor is essentially your teammate. Um and you use rules, skills, and sub agents to help with you. When you think about rules, it's essentially your standards that you must follow. Skills is kind of like a playbook or workflow that you can you can use. And lastly, sub agent are your specialists. Um, so

32:32 if you think about within a codebase, you probably want a back-end specialist, front-end specialist, and a database specialist. So these are your sub agents. If you look if you if we look more into it rules some key idea for it is you always follow these um it's like a coding standard or security policies or convention within your team that you always want to apply and these always want to apply rules could be um could be segmented into certain um files that's

33:08 being called. So let's actually go back into our IDE and open up one of our rules file. So you can see that I have a Java import MDC. MDC is markdown. C stands for cursor. It's essentially saying that always apply is actually false. You don't want to run this rule all the time because context is quite agrace resource. um you only want to run it when it's a Java file um to fix its imports if the import is not working well. Next one is skills. So skills um

33:45 is like a ondemand playbook. Um it's a you you saw in the in the demo that we create a start game uh skill that's essentially kind of like a command that the agent is able to run and also the humans are able to run it as well. You can manually evoke it as well. And the last one is sub aents as as I talked about earlier. It's like a specialized worker um in a autonomy and control stack. You can kind of see that sub agent has the most autonomy. Um, you

34:23 probably have less control over it. It want to oversee and call different skills and also it follows the rules that you created. So this is this is how you um this is how you use these three key concepts over here. Um let me just quickly check the terminal. It seems like um run this command once. What is it trying to run? That looks good to me. Um, so while while this is still running, um, there's some resources that I would like to show for you guys. Um, if you guys go to cursor.com/doccks,

35:02 this is um I've I've read multiple um documentations in my life, technical documents to my in my life and I have to say cursor has one of the best docs and one feature that I would like to highlight is ask AI. You can ask AI any questions. Let's say um give give me an overview of rules, skills, and sub agents and it would gp through and semantic search through our documentation give us a very good idea of what the question is and get provide us a very detailed response. Um, so for those of you that

35:42 want to learn more, you can use this as AI feature as well as if you come in here in the customiz customizing section, you have your rules, skills, and sub agents that you can read along. Okay, this is trying to This looks good to me. So it seems like uh um Cursorbot already finished its its um its robot right away. Uh whereas um GPT and also Opus, they're still working very hard on um trying to get the bot start running. Um meanwhile um John, do we have any

36:23 questions that I can answer? >> Yeah, definitely. So um a few ones here. One just on the rules, skills, and sub agents topic. Um, a few people asked about commands versus skills. Um, anything you want to share there or we can speak to there? >> Yeah, of course. So, for me, um, uh, correct me if I'm wrong, John. Um, commands are going to be, um, slowly, um, sunset in the future. Um, the reason why is because it's kind of like a, um, subpart of skills. So skills is um a specific um uh workflow that's in being

37:01 able to invoked by the agent as well as a human. Whereas command is like a workflow that can only be invoked by the human. So then why not just put that under skills? So that's why in this webinar I've only been preaching these three things instead of adding command as well.

37:21 >> Yeah. Awesome. Um, yeah, and I will say, uh, just in case anyone is worried there, we don't have concrete plans to roll out to remove commands in the immediate future. I think it's something we're considering just because, uh, skills are kind of a supererset of them, but that's something that we, uh, no concrete plans to phase those out or anything just yet. In case anyone loves their commands, we will do some user research around that before making any

37:42 big choices. >> Sounds good. Sounds good. Um, and let's let's actually see even though it's running, I think they're they're all running about test. Let's actually open up um a new game here and see if So, we have the cursor bot in here already. Oh, look at that. We have Opus.

38:04 Do we have GPT? Um not seeing GPT. So, it seems like GT is not there. Um Um that's fine. Uh we can go ahead and start this battle and see what's up. Oh, okay. Well, that was that was fairly fast. Um, it seems like cursor won by 100%. That's very interesting. That's if this is how we evaluate models that that would be a good good deal for us.

38:36 Um, uh, I think next thing I would like to talk about is, um, cloud agents. Um, or do we have do we have any, uh, questions that I should answer before we jump into cloud agents? Yeah, let's maybe take a a quick because we actually we have a ton of questions people are are loving which is great. Um I think before we shift into cloud agent it's probably worth um just clearing through a few of these. Um so one that was around uh as a Java project we have Java

39:02 coding guidelines and some linting rules. How do we make the agent follow along with these? I think you've talked about some of this already. Rules is the answer to that. Uh skills could be but mainly rules. Uh so that one I think we have we have generally covered there. Um and then uh >> sorry one one thing to add is um I know it's very daunting to first first of all when you when you realize that oh my god I need to create a rule what should I

39:26 type it there's there's so many things like it's very daunting to start what's very good is you can actually go to the chat and hit create create rule create sub agent create skill there's already a builtin um template within cursor that helps with a prompt that helps you um start creating this rule. And you can say um help fix lenting errors and then progressively it can help you keep building on top of these these rules. So it goes ahead and create some

40:03 rules. If you don't like it, you can change it within the same chat. Hey, I don't like XYZ. add ABC something like that and what's very cool is another resource for everyone is we actually open source our cursor team kit so we can say cursor team kit um it's a plugin um so we open source our sub aents our skills and also rules so we're talking about rules you can actually see that within here you'll be able to see within cursor team how do we define rules rules. Um, and contrary to

40:45 popular belief, our rules are actually very short. It's only eight lines long. Um, if you have rules that's over a 100 lines long, I would highly suggest you to maybe see if we can trim it down or perhaps have different rules um um in different files.

41:07 >> Um, we do have some other ones we could keep going through, but uh kind of fun if we have time to to touch briefly on kind of cloud agents. There was a couple people asking about like triggering things from Slack as well too. So, um I think cloud agent's a good beast to touch on there.

41:20 >> Perfect. Perfect. So, um going back to the slides real quick to look at the bird's eye view. Um we can see that we mainly focus on cursor desktop over here, Jet Brains IDE. Um here and also cursor CLI. We really want to touch base is the cloud agent. Um, have you? And so one one thing that was revolutionary personally for me was one day I was on uh Uber um super long um going through Baybridge and I realized that hey I have this super good idea on my project. I

41:56 really want to code this up but I don't have my laptop with me. Well guess what what you can do is you can go to cursor.com/agents. Choose a repository that you want to work on. Uh I want to work on demo. This is the Robo code that we were working on. Choose a branch that you want, which is main. And give it a command. So this one I probably want to say, um, hey, let's optimize waterbot. I can't believe it lost to firebot.

42:28 Just just something over there. Um, you can choose a model that you want. Um, so we used 5.4 earlier. Maybe we can use uh 5.3 uh because on terminal bench 5.3 is better. Uh you can actually choose what MCP server that you want to link it with. um what MCP server that you want to link it with here uh we can add slack into there and with uh with relating to John's question um we can say something like once you are done ping me on Slack and then the channel

43:06 you can say um a channel name of XYZ and then you can go ahead and kick that off in This spins up a brand new uh VM that has all your um all your information of what your GitHub repository look like. Uh it has all the system prompt that you put in the repo that you want to work on, the MCP that you want to put in and it has skill integrated as well. You can actually do slash commands with that skill. So slash um here get commit git push all these are actually get start game which is the

43:48 one that we created is repo specific and you'll be able to kick that off asynchronously. It is a very strong tool. You can kick off three five all at the same time and it's working on its own branch and present it to you. Normally these tasks actually takes longer. Um I would say at a minimum time it takes about 10 to 15 minutes. So maybe we'll we won't be able to see this particular output. However, I did come prepare for this webinar and I actually

44:22 kicked off a a a cloud agent earlier. And what's really cool about cloud agent is it not only provides you diff file of the PR that it created, it also gives you internally we call it artifacts. These artifacts are screenshots as well as screen recordings. So in this in this file here I have opus that run and say hey use your best judgment create a robot and then afterwards use this command this skill command start game and give me a demo out of it gives me a

45:00 a video of the bot that it created. Watch this. You can see that it's fighting fire. This purple one is Opus and it's just dancing around fire and it and it just demolished corner. It's very interesting how the cloud agent is able to code up its own bot and at the same time just gives me a video recording of what it built. And based off of that, hey, you know, I I I like this a lot. I can actually just go ahead and create a merge request which um let's go ahead and go to the demo re demo code and um

45:46 let's go to our poll request and right there we just create a poll request and we already saw the demo. We we can check the diffs if we really want to. we can check the diffs again to see if the code is good and if it's good we can just go ahead and merge the pull request which is phenomenal. Um it's it's one of the uh moments that's that's that's like wow factors for myself and um what Kusser really wants to do is integrate into software development life

46:22 cycle. So another thing that we are also uh another product that we also shipped out is bugbot. So within PR review you don't you can review it yourself. You can also kick it off to cursor to to to um to review too. All you have to do is just run bugbot run. cursor goes ahead and review and found one potential issue and say that hey the radar lock event is overridden try and fix it and if you want to fix it you can fix it in cloud where it just gives it a prompt you can

47:01 send that off or you can fix it in cursor and within cursor it just opens a project and then you'll be able to you'll be you'll be a to you'll be able to fix it. So, yeah, that's that's uh that's the bulk of the demo for for now. Um if there's any questions, um I'm happy to answer. >> Awesome. Yes. Thank you. Um uh sorry, just respond to a couple of them here. Um it's amazing. Okay, cool. There's a there's a bunch of questions that folks have. I'm trying to pull out some good

47:34 ones that we would uh make sure that we can go through. Um, I think one that people had questions about is um just whether or not we could integrate cloud agents with like GitHub actions or anything like that. So like when a PR is raised, do something. Uh, so there's one question that I think might be worth mentioning live. >> Oh yeah. Um, so that's a that's a very good question. So, one thing that um I actually didn't show you guys, let me open it is automations. What's very interesting is um within

48:13 within cursor we realize that hey there's a asynchronous workflow that we can do. We can use cloud agents. These cloud agents just work autonomously. All it needs another thing is a trigger. So what you can do is you can add a new automations call it Joseph automations that's fine you can add a trigger and these triggers could be poll requests could be draft open. So let's say today if I want to create a poll request let's say every time it merge I want it to be

48:50 like a document um document consolidation bot. you are a doc writing specialist. Every time when there's a PR that gets merged, you read through the code and op update the docs under slashdoccks. Just like that. Choose a model that you want. Um, in this case, it's just writing docs, so I probably don't need something too crazy. I'll just use Haiku 4.5. Um, you can add you can add different MCP tools. So every time when it runs, hey, send it to me, right? Send it to my private channel. Um, so I

49:31 know that it's gets it gets triggered. In this case, you can go ahead and push this. Um, oh, sorry, I forgot to select the the repository, select the Robbo code. So every time when there's a PR that gets merged, this automatically run. And you have many options over here. It doesn't just have PR merge. Um you have Slack integrations as well. So within the public channel if you type some something it can also trigger this automations. You have linear pager duty is a very uh very interesting one. Um, back in my days

50:06 when I was at Amazon, I get woken up at like 2 am, 3:00 am. If we had cursor, we'll be able to I I probably roll out of bed and see that, oh, the cloud agent is already working on it and already has it fixed. And I'll be able to review the PR manually and say, okay, this looks good. Ship it. And then within 10 20 minutes, it'll be it it will be finished fixing. So, it's it's very interesting.

50:32 >> Awesome. Okay. Um, I think maybe one more and then we'll we'll call it there. I think I know you mentioned this um earlier on uh but I think there were uh a few questions that came through around model selection. So maybe just worth kind of calling out. You I know you were mentioning it before, but just Joseph, how do you think about which model you want to pull in?

50:52 >> Yeah, that's that's a good call out. Let me go back to the slides. Maybe people people want to see the slides one more time. Um so one thing to tell people is um there's no right answer. There's no right or wrong answer. Go out there, play around with all the models. Um you'll develop your own personal preference. So um it's like asking people like what is the best um what is the best t-shirt to wear? Everyone has like different color preferences,

51:22 texture preferences. It's all over the place. So the best way for you to do to develop a good sense of understanding of what model to use is to go ahead and try them out. Um with that being said, I would also give you my personal opinion. So whenever it comes to planning high reasoning, I love to use opus as well as GPT 5.3 or 5.4. Um whenever there is doc writing, I love to use Gemini. Um um whenever there's a medium to to low task, meaning file edits is within five

52:02 files, I use composer 1.5 all the time. It's the fastest model. I can't stress how great it is. Um and then um one last thing is GPT models push backs a lot for me. Um, so whenever I want to audit a plan or find bugs, I love using GPT models and it scores the highest on terminal bench.

52:29 >> Awesome. All right, I think we can probably wrap it up there. I know we went through a bunch of info. Uh, Joseph, thank you for the awesome demo. For anyone whose questions we didn't get to, apologies there. I think uh thank you guys all for all the questions, all the engagement. We will make sure that we share out this recording to uh to everybody uh who signed on for the webinar. So we will make sure you guys get that afterwards and uh we can

52:52 include the slides as well in there too. So thank you all. Joseph, thank you again and everybody go try out the new Jet Brains ACP integration that we have with Cursor. Excited to hear what you all think. >> Sounds good. Thank you guys. See you guys.