Cloud Agents

0:00 Hey everyone. My name is Emily and I'm an engineer on the growth team here at Cursor. And today we will be presenting about cloud agents, which is agents that run in the cloud and they keep going when laptops are actually closed. And they let you do long-running tasks in parallel across multiple agents. And with the harness that we set up they're actually able to self-test, build artifacts, and deploy and generate kind of previews for you to come back and evaluate. You're able to actually

0:28 run them on a schedule with our new automations feature, not just on demand. And they can run from minutes, hours to days and report back with the real results that you can verify. Some key things that we've developed for cloud agents, one is a really new harness that actually lets you run cloud agents for long tasks for maybe even weeks at a time.

0:49 We had Wilson, one of our researchers, go ahead and actually develop a browser across thousands of commits. Um and it was able to actually simulate almost Google Chrome functionality. We also gave cloud agents the ability to um have [clears throat] artifacts. So they can do computer use, click around in the browser, um and actually simulate the process that a human does when you're verifying code and testing the output through the visual and actual final output. Um you're actually also able to launch

1:20 agent swarms. So you can launch a bunch of different sub-agents that run in parallel to do the task and report back to the orchestrator. And then finally we built a lot of extensibility in mind. So you can actually build a bunch of custom skills, plugins, and hooks so that the agents work best for your internal environment. Um a few key tasks here. Um one is the sub-agents that I've been mentioned is that you're able to actually, let's say you're doing unit tests or writing your larger factor, you're able to have all

1:49 of these agents working in parallel to get the job done once you figure out the right plan for them to implement. Um and then with artifacts you're able to actually get videos or even screenshots as an output. Um so you can quickly go and verify the results. Um and as I mentioned also we have our new automations feature. We won't be diving too closely into that today. We'll have a separate sessions. But you're actually able to specify agents that run either

2:12 on specific triggers, so a Slack message, um a GitHub PR, or on a cron job to actually do these continuous tasks that should always be running. And one thing I did want to highlight is we internally use cloud agents a lot. Um ever since kind of that artifacts release back in early January, um our internal usage has skyrocketed and now about a third of our PRs are actually created and merged by cloud agents. Um so you can kind of see this is not something we've just built, but this is something we heavily use

2:39 internally and I'd say for me I'd say 70% probably of my PRs are actually by cloud agents. Um and yeah, I think a little bit more on how Cursor uses Cursor, like we don't agents don't just write our code. Um it reviews our code, it tests it, it fixes it, and it also deploys it. Um we kind of made it as easy as possible for us to quickly iterate and launch code and new features to all of you. Um and kind of a vision for how we see coding agents evolving. One is that they work best when they

3:10 have knowledge to your internal tools and systems. So being able to surface the right MCPs that interact with your um knowledge base, being able to write skills that provide that relevant context for how best practices and how all the um internal systems interact is actually critical for getting the best results that you want. Um and so we launched a plugins marketplace um a few weeks ago to enable that to be as seamless as possible. And last week we

3:33 launched two marketplaces so you can actually share these um shared knowledge bases across different team members. Um and then agents are going to the cloud. I think less and less time will be hands-on keyboard and you will be kind of orchestrating a bunch of um PMing a bunch of agents running in parallel. Um so if you have them across different repos working on parallelized tasks, um your kind of will be job is to working on making sure they're working

3:57 on the right thing, constantly iterating and looking at the outputs they're giving. Um and then the bottleneck will become more more on this review and code quality, which is something that we're focusing on um a lot as well. It's how do we make sure that with the volume of code being generated that it's not resulting in any um vulnerabilities or bugs that are coming out. So we have kind of a code review product called BugBot that we use a lot. Um and then

4:21 with automations you can actually also build in a bunch of custom flows that do this verification as well. And of course there has to be controls in place. You don't want a lot of agents running um rogue out there in the wild. And so making sure we have the visibility, flexibility, and the control is something that is always top of mind for us.

4:40 Um so that's a high level a little bit about cloud agents, but a demo is worth a thousand words so I'll hand it off to Emre that to show the product connection. Amazing. Thanks Emily. And again reminder folks, if you have questions, please just put it in the Q&A. Emily and Florian will try and get to them um while I'm talking. Um cool. So thank you Emily for talking uh or introducing introducing cloud agents. I want to before I kind of jump into the demo here explain why someone would want to use cloud agents. Um cuz I

5:08 think we get that question a lot is like why use cloud agents versus local agents? The main reason is um kind of uh I want to say async or multitasking or kind of orchestrating agents here. Basically how I use cloud agents is I have a list of tasks I want to get done. Um maybe it's the end of the day. I like to close my laptop, maybe go for a walk, or even just get other things done on my local machine. Firing off cloud agents has it run on remote VMs, meaning that

5:36 it's not going to be constrained by my machine, and it's also not going to, you know, take up any RAM or uh extra space that I might not be able to afford. So what's great about cloud agents is I can now run as many agents as I want on a external infrastructure, on the Cursor's infrastructure, and that will be able or that won't affect my own personal uh machine or things I'm doing locally. So that's kind of one big reason. Another is um and I'll show

6:01 here, cloud agents are just really good at self-verifying their work. So I'm going to share my screen and we're going to go through this a couple ways. I'm first going to show the docs so that folks can see if they want more information how to learn about cloud agents. Um we have a really great kind of summary here about why you want to use cloud agents. Um but I think more importantly than that, if folks are interested about uh security and specific settings, let's say you're using Docker or you have a specific

6:27 configuration, um we have really great documentation around it. I think more importantly than that, um cloud agents are really, really great at just doing work um that is something that maybe you thought you wouldn't be able to have to have time to do. So it almost feels like an extra friend or colleague that is picking up work for you that you might not have thought you'd be able to do.

6:49 Cool. Okay, so again, go to these docs. Um Florian will put this URL in the chat if you want to kind of enhance your learning as I'm speaking, but let's dive right into how cloud agents work. So I'm going to show two things. I'm going to show how to uh start a cloud agent um in like a brand new repository. And then I'm also going to show what a cloud agent run looks like and all the capabilities that you get with cloud agents. So the first thing I'm going to

7:11 do is I'm going to go to cursor.com/onboard. And what you're going to see here is I can actually select a repository of my choice. So I'm actually going to select my personal website cuz it's kind of small so I'm hoping that we can get it fully built out by the end of this demo.

7:25 Um I don't have any environment variables or secrets that I want to put in here, but I can if I want. Again, if you have questions about how these are stored, feel free to check out our cloud agent security page. We talk a lot about how we do kind of the cloning of the repo and the snapshots and secret protection and all of that. Um so I'm going to go ahead and just go ahead and hit start for free. It now is going to set up the development environment for this codebase. Um a couple things you'll see

7:49 here. The first is that you have this really great setup where you have the repository uh development environment being set up on the left, um all the dependencies being installed, and then you have um what we call the cloud agent UI on the right. So you'll see here it has kind of a list of things that the cloud agent is going to do. Um it also has your list of secrets if you want to see those. Um and then also it has your Git UI. So as uh you make changes in the

8:13 cloud, as you ask Cursor to implement things, you'll start to see uh the diff showing up, seeing uh meaning that you still get transparency into the code. Cursor's uh a big believer that you should always be able to see the code if you want to. And so you'll be able to see your diff here. Um you can, you know, see your commits if you've committed stuff already. And what I'll also show as well in a previous run is you can also see CI. So if you have CI

8:37 checks running, you would also be able to see that here. Um what's great about cloud agents is they automatically check out a branch for you. So here I have a branch with the development environment being set up. Um and I'm able to kind of check that branch out locally on my machine if I want as well. As you're seeing here, cloud agents can start sub-agents. Um that's a really cool feature. Um and what that means is it's able to work in really, really large codebases. So

9:02 uh I'm going to show you uh a cloud agent working in Grafana, which is I think something like 2 million plus lines, 30,000 files, and it's just a testament that Cursor is great for large repos. If you want to upload your mono repo or whatever to your cloud agent, you would be able to.

9:19 And then the last thing I want to show is uh you and we'll see this in action is Cursor actually has access to its own computer and mouse. And what that means is that you'll be able to kind of uh test things manually for you. Um looks like it's kind of not connecting to the desktop today, but I'll show you a previous um a run of this and you'll see what that looks like. And then the last thing is the terminal. So you still have access to terminal to see any commands that Cursor's running, um if it's running a

9:44 local server or anything like that will be all in the terminal here. So that's kind of a little bit of a tour of the cloud agent UI. Um pretty straightforward, really great uh again to visualize what's going on. As you can see here, it's uh running uh based on what I've given it in my repository a couple of scripts um and it's going to try and understand what's going on here.

10:06 As you can see, it's saying this is a personal website, all of that, and then now it's going to install all these dependencies for me. So, pretty straightforward, a great way to kind of set up your environment. Please check it out. Um it's really uh a a really like fast and also just easy way to get started in the web. Um and the last thing I will say is that this integrates with both GitHub and GitLab. So, if you are not on GitHub um and you're on GitLab instead, feel free to use Cloud Agents. Um we have Bitbucket

10:33 support coming soon for all you folks on Bitbucket. Um but feel free to try it out on your own environment for now and we will uh we'll talk more about what that looks like. Cool. So, this is the setup environment. Um now I'm going to go into like what a Cloud Agent run looks like. So, there's multiple ways you can fire off a Cloud Agent. Um the first way that oftentimes I do is I do it from the Cursor IDE itself.

10:56 So, as you can see, um I asked it to build a plan. I'm going to ask it to build this new dark mode plan for my Grafana repository. Um I'm going to say invisible. Um and what I'm going to do is I can actually build a plan locally and then have Cursor implement the plan in the cloud. And I love doing that because it's a really great way of kind of working with Cursor to understand um the plan and like what my strategy is. This is my Devil's Advocate sub-agent

11:23 running actually that helps me kind of challenge what's going on in my design questions. Um and then I can run it uh in the cloud once it's done building the plan. So, that's one way to do it. You can also just run um agents in the cloud locally yourself. Um so, you can just go from local to cloud and then just fire off an agent without having to build a plan and then that will automatically start running here as well. So, I can do that. I can say change the font of the

11:47 Grafana main header to Times New Roman. And what you'll see is that it will start all the environment set up, it'll allocate all the resources, and what's cool is you'll actually now see that agent running on uh over here as well. There it is. The Grafana main header font change is now running. So, this is like a great way to kind of sync between your local and your cloud environment.

12:09 They're always there um and everything will always be synced. So, that's one way you can launch it. The other way to launch a Cloud Agent is actually from Slack or Linear. So, you can actually comment on a Slack message um or a Slack thread and say like, "Hey, @Cursor, can you solve this for me?" Um that's due to our amazing Slack integration. A lot of PMs use this feature when bugs come come in and they want to be able to fix it easily. So, that's one way you can do it

12:33 as well. And then Linear is another way. Um you can do @Linear in any ticket and it will launch a Cloud Agent and be able to solve your bug in the cloud. Um the last thing I want to show with re- regards on how to launch a Cloud Agent is actually uh mobile. So, I'm actually going to bring up my phone here. Let me go ahead and open it um so you all can see.

12:54 It's connecting. And what you're going to see is I actually have cursor.com/agents up on my phone. And as you can see, I can see all the agents that I just launched. So, I have my dev environment setting up on my personal website and then I have the header font change on the Grafana app. So, everything is kind of in one place. I'm able to see like what repository is working with what agent. So, it's a very easy way to kind of do multiple multiple uh like agent uh requests across different repositories,

13:23 which is really neat. Um so, I know a lot of people are asking about like cross-repository work. Um this is kind of my favorite way to work in like both my website and my Grafana uh fork and like the Cursor website that I'm making changes to all at one time. And you know, I can click in, I can launch another agent from mobile here if I want. Um I can even go inside here and check and see how my agents are doing. As you can see, uh this is an example of

13:48 uh Cursor using the computer. So, we'll we'll take a look here and see what it's doing. Um but again, using using your mobile phone for launching Cloud Agents is really great. It's very seamless um and we're really excited to have this feature up. Um Cool. Okay, so that's mobile. I'll go ahead and just minimize this for now. Um I'm going to go back to our plan. Oh, looks like the plan is asking me more questions. That's fine. Um and then I want to show you all the uh build in cloud feature cuz I think that's a

14:16 little bit hidden in plans and I want to make sure that everyone sees it. Um in the meantime, I'm going to minimize this and I'm going to go back to where it was setting up my um my website. Cool. So, as you can see, it went through all the to-dos here and now it's actually doing the uh computer use, meaning that it's seeing how it actually looks like to start the server and uh run my website on um Cursor's Cloud Agent computer. As you can see, it

14:43 checks for uh API responses, so it makes sure that everything is returning 200. Um but it can sometimes take longer than, you know, maybe it would locally. We keep a little like note here saying that Cursor setting up a development environment from scratch can take from 5 to 30 minutes. So, just be aware of that, but remember that you can always launch as many agents as you want at once. So, this should not be a blocker. You can go ahead and just go and check

15:06 on your other agents in the meantime. Cool. So, this font change is working well. Um now while kind of these are both running, I want to show you all what a completed Cloud Agent run looks like and how you can see uh demos and artifacts like what Emily showed. So, I'm going to go actually to a Let's go to this guy. This feature that was built. Cool. So, again, I was working in uh Grafana, so my Grafana fork, and I wanted to implement a new feature. So, a

15:33 couple cool things here. The first is that Cloud Agents in Cursor have access to MCP servers and this is huge because uh to my knowledge, no other uh AI tool Cloud Agent right now can use MCP servers um in the way that Cursor can. So, all I asked it to do was implement this ticket, graph-59. It immediately was able to understand that this is a Jira ticket. It got all the full details from the Jira issue and then it made a plan. It explored patterns using

16:01 sub-agents and then it read the key key key files that it needed to modify. What's really cool about Cloud Agents and just about um Cursor generally is it is always able to parallelize work with sub-agents even in the cloud and that makes both the speed and the comprehensibility of what it's able to do really powerful. So, I think that Cursor is the best way to work with large code bases and, you know, and be able to make changes in, you know, that

16:31 that doesn't take like hours. So, I love working uh with Grafana in Cursor because it's just super super fast. So, I asked it to implement this ticket, graph-59, which is basically a way to view the feature flags in Grafana um and then just view it on uh a separate page. So, if you want to see the ticket, I can go ahead and just load that up right now. Um graph-59, oops, 59.

16:54 Um Oh, sorry. Graph-59 Jira. Hold on. Um This is probably where There we go. Here's my ticket. Um so, as you can see, I have details about what the ticket does. Um Cursor reads all these details and then goes ahead and implements them step by step in the Cloud Agent. Um so, I basically said, "Hey, like feature flags right now are not visible in Grafana. Can you make a whole page that shows all the feature flags and allows me to toggle them on and off?" So, that's a pretty big feature and this is Cursor actually

17:28 sending me a walk-through video after it's finished implementing the whole feature on how it did. And this is really phenomenal. This is like what you would expect kind of a colleague or a co-worker to do after they finish implementing a feature, they add a video to your PR being like, "Hey, this is proof that it works." Um but Cursor does this all on its own. And as you can see, I didn't really give any instructions on how to test it. I gave it a scope, I

17:53 gave it acceptance criteria, I gave it some notes, but Cursor was able to figure out on its own how to test a feature like this. So, it was able to kind of see how to search for stuff in the search bar. Um it was able to understand like how to toggle flags on and off. We really think that these artifacts are the future of what coding is going to be like where these agents will be able to go off, make changes, and then come back and be like, "Hey,

18:17 here's proof that this change was made and I've made a video for you showing that." So, it's really really neat and very cool. Um You can save these videos if you want and share them with your team. You can get a link to the video. You can also just ask Cursor to upload these videos to your PR if you want. I've done that a couple times if I want um if I want to have those videos saved. And then we add our lovely little Cursor logo at the end so you show that this is Cursor made. So, these videos are really

18:47 neat and this is kind of my favorite part of Cloud Agents are these artifacts that Cursor creates. As you can see here, it also takes screenshots if you want not a video and just want to see like what each area looks like. As you can see, there's a screenshot of it searching in the search bar making sure search works. Um there's a screenshot of it checking the other uh pages that it created. So, it was able to categorize all the feature flags by experimental

19:11 versus preview and then it was able to filter it by there. Um so, this is like truly remarkable. This is a big feature. I mean, I was a software engineer for 6 years before this. I think implementing a feature like this probably would have taken me weeks um and Cursor was able to do it in an hour. Um so, that is just remarkable. In addition to the manual testing that it does in the browser, it also does a lot of uh testing locally. So, it's able to run scripts.

19:38 Um again, I didn't give it any information on how to test it. It was able to figure that out all on its own. Um and it so it made sure that not only did it pass the test manually and it looks good, but it also passed all the unit tests um that it was able to create and run for itself. Um as I mentioned, there are CI checks that happen and this is just because this is uh a forked repo of Grafana. So, I get all the CI checks that would be on the original Grafana repo. Um you can actually have Cursor access these checks

20:07 and be like, "Hey, why is this check failing?" and it will rerun CI and be able to understand and figure out what the checks are. Um this is something that actually is a great use case for automations. So, we actually have a great automation template for uh fixing failing CI checks that we'll probably talk about in a later session. But again, really cool way for Cursor to see CI and understand what's going on there. And then you can also see your commits,

20:30 so you can see what what changes were made, what changes Cursor decided to make. Um you can check out those commits if you want more granular detail there. Um but it's a really cool way as well of making sure that you're not just creating massive PRs. Cursor is organizing them in a way that makes sense. So again, this is like a what a final cloud agent run will look like. Um and I wanted to show you all kind of the beauty of it because I haven't seen this in any other tool. Um this kind of just

20:56 blew my mind when we first tested out internally. So, I'm excited for you all to try it because it really is quite neat. Um as you can see here, it looks like our dev environment is still running. Um a couple of things that I want to mark here. Oh, there you go. There's our little walk-through video that we made for my website. So, as you can see, it's going to go, it's going to go to my localhost, have my server running, and then it's going to show off

21:18 my website, hopefully. Let's see. Looks like it's refreshing, making sure that everything looks good. Um and there it is. And there's my website. So, this is just a really cool example of, you know, and this honestly didn't take that long.

21:31 Oh, I love this. It's testing and making sure that all my blog posts work. That's very cute. Um so, it's testing stuff like that. It's making sure that everything, all the other uh links on my website work. I have a little connections page on my website when I used to make my own connections board.

21:44 Um so, it's doing a really good job of that. And then it's also navigating to uh Oh, no. This is just a screenshot. Amazing. But as you can see, it does a little summary. Um so, it says, "This is what I did. This is all the things that I tested." And then what you can do is obviously you don't want to uh rebuild this environment every single time. So, when I actually actually save this environment to my team so that new agents can start from that current

22:08 snapshot and they don't need to always rebuild the environment from scratch each time. So, we're totally aware that, you know, the cloud agents can take time, especially for something like Grafana, it probably takes a while. So, definitely save your environment to your team so that you can start from that same snapshot. >> [gasps] >> Cool. Awesome. So, this looks great. I can now go ahead and just start uh talking to the agent and making changes. Um

22:30 I like to actually go to this uh UI for new changes. As you can see, I can choose what repository I'm working in, so I can change repositories really easy. But then I can also choose what model I'm using. So, I think this is a good thing to highlight is we have all our greatest models and this is a model we released as of the last, I think, 10 minutes or maybe 20 minutes, Composer 2 is officially out. Um please give it a shot. It is a great great model um and

22:58 it's much much cheaper than a lot of these frontier models at the same intelligence level. Um but if you want to kind of try out other models, um this is a great way to do it. I really like Codex 53. I think it's a really great model for how inexpensive it is. So, if cost is something that's important to you, definitely try out Codex 53. And then as always, you can try out multiple models and see how they work. Um we offer that setting in cloud agents as

23:22 well if you want to compare models. The last thing, or actually two last things I want to talk about, is the MCP servers. So, I mentioned it earlier when you saw that I just asked it to implement a ticket, it was able to access my Atlassian MCP server, but you can actually add whatever MCP server you want to your cloud agent. So, any MCP server that you have locally, you should be able to add. I really like um the DataDog MCP server. I think it's really

23:46 great um in like identifying uh you know, performance issues or any bugs that have come up. A lot of folks also really like um Where is it? Oh, there it is. Um I guess the Slack MCP server and Atlassian MCP server are both kind of, you know, par for the course. A lot of folks have used them. But again, Linear is a great one if you want to create more linear tickets. Um and then we have a a lot of other cool ones that people have enjoyed. Um I think one that specifically I like is

24:15 doing the um the integration with TLDraw and with actually the Excalidraw MCP server, which I think I would just add here to my custom. Um mainly because I love uh Cursor to create diagrams for me and I do think that TLDraw and Excalidraw do a great job with that. Um I also have the Granola MCP server turned on and I use that sometimes to uh take notes from meetings and use that with Cursor to generate action items and feature requests. So, there's a lot of great things you can do with MCP servers and

24:42 cloud agents and it's it's really quite powerful. And then the last thing I'll show, this is actually in beta right now. So, I think maybe some folks will have access to it, I'm not sure. But you can actually have a long-running task, which means you can actually ask Cursor to work on a task um that can take maybe hours. Uh you can go for as long as you want if you do grind until done. So, um as Emily mentioned, we had one of our engineers, Wilson, build a browser, like a full browser, from

25:09 scratch using grind mode. Um but yeah, I've seen people use this for like really, really uh tough migrations. Um I know someone who tried to migrate from Poetry to UV doing this and they were able to do it. Um same from like SQL upgrades from like version 4 to 5, um you can do that with uh with something like grind mode. So, check that out. Last thing is we offer the same voice mode and images that we do uh locally with Cursor. Um so, sorry, with the

25:36 cloud agent. So, that is also um available for you if you want. Um but yeah, this is kind of a tour of cloud agents. As uh Emily said, we'll do a specific um deep dive into automations next week, I believe, um where it's actually going to be cloud agents on a schedule and on a trigger. So, you'll see here there's a ton of different templates that we offer. But if you kind of look inside one, so here's like a cleanup feature flags cloud agent, um you have one that actually you can schedule on a trigger

26:06 um that acts as a feature flag cleanup kind of bot for you. Um and then you can, you know, connect it to whatever MCP server that you want. So, really quite neat. Um I know uh I'm about time. Are there any questions that I could tackle before I end? I think there's two core groups of questions I wanted to highlight. Um one is people are saying they have a separate like front-end and back-end repo. Um cloud currently doesn't support multi-repo workspaces. Is that correct?

26:36 Correct. Yeah, so if you have two repos, so you can you can work on them separately, but you can't work on them together if that makes sense. But that is coming very, very soon. Yes. And then the other common question is a lot of people work at enterprises and they want to run cloud agents in their own cloud. Um that is on our road mapping road mapping coming very soon.

26:54 So, um stay tuned for that. Yep. And you'll see that actually right here, this use private workers, it's in beta right now. We're testing it out with a couple customers. So, that will be your your gateway to running cloud on your own infrastructure. Mhm. Yeah, and if you wanted to get connected, feel free to email me Emily@cursor.com. And I know we are at time, but yes, highly encourage you to all check out our new Composer 2 model. We're all really excited about it. And yes, this recording will be shared after the call.

27:19 So, thanks everyone for joining.