Cursor's Hidden AI Agents Feature Could 10x Your Coding Speed

Cursor background agents just showed up. They released it. They pulled it back. You got to get on the list and try to join it. Let's talk about it. So, I made a video about this yesterday. Yeah, yesterday for our community and I wanted to share it with you. So, they have RCP that showed its pretty little face right here in the beta settings. If you're not on beta access, make sure you turn that on. And you can do that by going into your cursor and going to beta features and just make sure that this is on early access. So what did it tell us? I'm going to pull up a more clear picture of this. So I think it has to do with background agents, but there is a protocol as well and there's a good blog post people asking about it. So it is not remote procedural call, it's RCP, remote cursor protocol. And so this is a bit of speculation. We're going to cover this and then we'll get to what I think. But essentially, it allows you to turn it on and then you can access your cursor instance through an external party. That's actually a huge game changer. cursor has not opened up much and it's going to be similar in my best guess to MCP where essentially it is the USB port into accessing all of the behavior that they determine that they want to allow external services to manipulate and so it's set to be on a per project basis and you can enable it in there you can configure a port and an authentication layer where you can have an optional access token and then you can persist that configuration to the project. Then you restart cursor. Some of the use cases would be custom ID extensions and plugins. You can think of, I don't know, client. I always talk about client because they're always ahead of the game. But in this case, you have an MCP server library. Very, very nice to have that in there. And you can't do that with cursor. You can't really attach much to it. You could have automated code review bots. I think that's a pretty interesting one. I like the idea of CI/CD integration, document generators, all those things. And it must connect with the background agents, right? So, let's take a look at this. Let's turn it on dark mode. And with background agents, you can spawn off async agents that can edit and run your code in a remote environment. Very, very cool. I don't have access to it. Give me access. But basically with command apostrophe you can submit a prompt and then hit that to view the status and enter the machine the agent is running in. So they had shadow background jobs before and I don't know if you guys knew that but it could be like pretty CPU intensive. So I'm curious to see how this is going to run but you can also join their discord. People are talking about it. When you first try to use background agents in a new repo you'll be asked to set up their machine. Think of it as an onboarding a new employee. This is exactly like a video that I just made on my daily channel. This is my main channel if you want to check out my daily channel. I just covered something similar where Brett Taylor, the chair of OpenAI. So that this agent shift will change the way that we sell software and buy software. So if you're a developer, which you most likely are, we need to be thinking about that. That is an advantage not a disadvantage because all the companies that are out there selling per seat they're facing the inventor's dilemma where they don't necessarily have the incentives to try to make change but they have to. So we can think of this in our case if we're building aic software that we are creating outcomes and we're selling outcomes and those can be priced. They can be very very expensive and beneficial for both parties if it's a harder problem to solve. But yeah, they won't be productive if every time you ask them to do a task, they need to clone your repo and install all dependencies from scratch. If your repo is complex, you should be able to expect you should expect to spend an hour getting the setup correct. crazy. But this reminds me of what I was doing two weeks ago where I would open up a terminal and I was testing out Cloud Code. If you're not familiar, Cloud Code is doing 30% off for developers until July 31st. You can go to your privacy settings on Enthropic Workbench and just opt into it. But what I was doing is I basically had I tried to get to six different work trees for a repo and had cloud code running in all of them. I got to three, but it was a little bit hectic. But it's that same kind of idea. Cloud code's going after more like raw CLI CLI style thing where it might not be for everybody. But yeah, to ensure best agent performance, make sure you set up the machine fully. Make sure all lentters run and that the agent can run your app and test. If you don't set up the machine properly, the agent will get distracted by setup. The machine setup is defined by a cursor environment JSON file. This reminds me a whole lot of docker compose. So if we open this up, this is their schema. And if you want to, that is the URL. Let's make this a little bit bigger for you guys. But this defines all the things that you can do. So pretty cool. You can see docker file container. You can see snapshot container. And what else is interesting in here? Yeah, basically all the Docker file related things. It's kind of like if you've used Ansible, it's it's basically just like a config file or even like Superbase is the config tunnel. But yeah, so which can either live commit in your repo or be stored privately only for your user. They recommend that you're going to do live commits and the setup flow will guide you through a proper environment JSON which we just covered. You're going to be asked to configure your GitHub base environment for the machine. Maintenance commands that should be run to keep the machine up to date. So, I'm wondering is this this? Yeah. So, you're basically running a VM and it's sick to think about. Very cool. So, background agents currently clone from your GitHub repo. They also do work on a separate branch and push up to your repo. This reminds me a lot of Devon. Very cool. Reminds me a lot of what we're seeing out of augment. haven't gotten their remote agent access, but they're all meeting on this one spot. I think this is where this goes where we have a really crappy first look at what agent management looks like today, which is right here, where you just have these little things and they're technically agents, but like custom modes are tough, the tooling's tough. You got to click through all these boxes. Like the fact that is this small is clearly hasn't been prioritized, but they're testing, right? And you have advanced options, you have autofix. The only one that I've used is memory bank and it has this giant prompt in it but not that great. So I'm excited for better management on the UIUX side of this and it means that you grant read write privileges to your repo. Yes. So I can imagine where I'm just going to have feature branches with these things running. You define the best PRD possible. You do that through questioning. You do that through sharpening your product strategy skills and getting more technical and using different prompts. Like one of the prompts that I use, if you guys want it, is one that I've crafted from being a PM for like ever, but it basically gets you to poke holes in your idea and come up with open questions. So, it's more of a conversation. It is a mental jousting session, if you will, which is really helpful. So, PRD skills going through the roof on value when it comes to agentic stuff. The base environment will define the hard drive of the machine that it'll run on. Yeah. Yeah. Yeah. It's going to be running on a YUbuntu image. Cool. You'll be asked to manually install all dependencies in your repo. Okay. This is something that I've also spoken about where I got a Net Cup server out of Germany. It runs on Debian and it's $12 a month, which in Digital Ocean dollars is like 160 for how beefy the thing is. And I've wanted to build an agents server for both growth agents, but also just things that I need for business. And I just haven't gotten around to it because I've been building this community and building my own SAS products. But this is that, right? It's a dedicated machine that offloads the CPU that lets it run on its own. And I love that it includes like these maintenance commands and all these things. I think it'll scare away a lot of people that are less technical, that don't have any DevOps. I don't understand why people are afraid of learning Docker. It's very simple. Just go read. They have extensive docs, but I think that's like where you start to dip your toe in the water with these concepts. You'll be asked to manually install your needs. Yep. So, you're using pseudo. If you're using Debian, it' be yum. You do an aptget. If you're not familiar, apt get is basically like a think of it as like a package manager. It is a package manager. So, in Rust, it'd be cargo. In JavaScript, it' be npm or pmppm. You got it. And let's see. Then take a snapshot. Taking a snapshot persists the disk state. The declarative setup uses a Docker file to define. Cool. Similar to how dev containers work. If you know what a dev container is, then that'll be helpful for you. Should not copy in your project, instead be cloned. Yes. And so we'll start from the base, run the install, yada yada yada. After running install, the machine started. We run the start command followed by starting any terminals. So cool. Start command can often be skipped. One common case where you want to use it is if your dev environment relies on docker. That's me. In which case you would want to pseudo service docker start. Okay. Terminals are meant for your app code. Yep. Terminals will run in a teamox session that is available to both you and the agent. Very cool. So yeah, I imagine you'd have as many VMs as your computer can take, which they're not even running locally. If you can cloud host them, then this kind of sky's the limit. That's the environment spec. What else? Oh, that's what we covered earlier. Oops. Where did max mode go? Max mode. Max mode. Only max mode compatible models are available to use. So it will be a little bit pricier eventually. We'll start charging for the dev environment compute. Cool. May also start. Okay. Interesting. So they're covering dev environment compute. Okay. Background agent has a bigger surface area of attacks compared to existing features. Specifically, you'll need to grant readwrite privileges. Yeah. Duh. Run inside AWS. Security. Security. Auto runs all commands. Yes. Yellow mode on perma exfiltrate the code. Okay, cool. Yeah, so I'm really excited about this. I didn't even know this notepads thing still in beta. I don't know. Forget it. But yeah, this is all really exciting and I want to get it because it's going to be dope. The way that I would imagine this going is in your files like I basically as a project goes I try to bas I try to be realistic on my level of competency and whatever the stack is. So in this case building out echo which automates uh the problem it solves for me is it automates all the grunt work around content creation. I like this part. I don't like the stuff around it. And so this has fast API as the back end with websocket to connect to tanstack front end. And then it has a superbase thing for basically authentication, managing user data, DB, yada yada yada, Google sign on. But I was not using much AI stuff to begin with. And that same thing will apply as AI gets better and better and better. So with background agents coming out or remote agents in the case of augment, you basically just save a bunch of tokens and time if you have competency. So you always want to pay down your ignorance tax upfront. Otherwise, that will compound and then you just run in circles. You can eventually prompt your way through it, but I think it on the on average, I feel like I end up spending the same amount of time if I'm just prompting and grinding my way through yapping. And I should have just ended up like going and reading documentation. Slippery slope for everybody, especially newbies, because the amount of dopamine that you can get for the current day Hello World is much 10x that of like a blank white page. you get like the fancy frame animations out of Bolt, but you need to be realistic and be like, hey, I actually want to be confident in these areas. And then as you start to learn the stacks, it took me l like realistically 10 days of just like grinding on fast API and tan stack start to get it to click and I was not using much agent stuff. But then as I got it then it's okay cool let's at least add in memory bank. Then I had memory bank as an instance within each one of these. I had actually cloned out my cursor rules. So there was a memory bank for each one of the different areas of my monor repo because this is a polyglot meaning multiple language monor repo meaning all of them are in one spot. It needed specifications. I needed memory banks in each one. I'm curious to see how background agents will take in context. I hope that it's similar to what we saw with ader or if you're not familiar wow I got someone to refer if you're not familiar with CLI tool pioneered the space. You can't see this. Let me see. But if I typed in oops hater and it's going to open this thing up. Not right now. Not right now. Not right now. Oh man. But basically like I can like one of the commands in here would be readon. So I could do for slash readonly and then I could have that be as a config for each one of the chats moving forward. In this case they only had architect and they had uh basically like code mode. So they have those two. You see the same thing with Klein where they have plan and they have act. I think that that's going to be a big important one which is like how will they traverse the gap between this little dinker right here and some sort of interface where it's like very easy the jobs to be done will be I need to manage context. I need to manage tools. I need to manage readonly versus writes. I need to manage how you deal with hitting a wall. Are we going to get a notification when the background agent has burnt through $50 in credits? Are we going to have spending limits on that? How do we know when something's going ary? Do we get a notification like you're seeing with Zed? And that's honestly a limitation of even how they've built this. This is a fork of VS Code. So, what's that going to look like? It should be net new. like they shouldn't have any of those issues that you would run into when you are a fork of VS Code. But I'm excited to see how it goes. If you guys want the prompt that I mentioned for the PRD, you can just comment PRD. If you want to join our teaming group of smart peeps, you can join our community. I typically end up picking like one story or two stories or a few stories out of the dozens that I put inside of our Discord. And then we also have a bunch of stuff in our school, which is cool, too. So, you can check that out. Link is in the description. If you found this video helpful, make sure you subscribe to the channel because I post these all the time. And make sure you like the video. I'll wait ever so patiently. Did you like it yet? Thank you. You did a great deed today. Good job.