This AI Setup Will Outperform Engineering Teams (Cursor Background Agent + Claude 4 + GCP)

using background agents to build agents. Things are getting really meta. I want to talk about that and I want to talk about the loops that I've been working on. It's been a minute since I've posted. I'm Parker X. I let tech for a startup that sold for 23 million bucks and I've been trying to figure out how do I leverage AI to get the most out of it for marketing, coding, and business. So, this is a daily upload channel. It hasn't been very daily lately because I've been working. I've been trying to build systems that will make the content better and the entire flow better. So, first of all, this is what the cursor sort of agents look like. Now, I was confused because I was not using them till last night, but your boy's been shipping. We have a new codebase. This is the AI community platform monor repo web app with discord ooth stripe subscriptions discord bot and shared packages for AI SAS builders community. Cool. So this is the repository that I've been working on and I think every commit was from cloud 4 almost maybe not the initial one was doing some scaffolding but the job is so different now like claw floor really it's just good it's very good and it's two cents per request on cursor And I was like hesitant to even mention that on video cuz I'm like, "Oh man, if they change it." But I took a screenshot of it and I threw it in Excaliraw because I thought it was just bananas where it's oh cool like 69 discounted for Cloud 4 request for $138. Crazy. And yeah, so I was I've also just on like the systems part I've been trying to figure out like what's a better way of laying this out. Do I just have these little billboards which are things about different products that I want to talk about? Maybe have an area for a catchy title at the top. But yeah, so Cloud 4 absolutely ripping and the way that you set up the cursor background agents is it's a little buggy, but you basically you hit command and oh command and semicolon and it'll pop open the pane. This is very reminiscent of cursor composer if you use that. Shout out to the OG cursor users. But I was throwing really simple tasks at it. It only allows for you to use Claude 4 Max, so the Max models. And what I noticed is this thing just will never finish loading. No clue why. I've tried restarting. I've tried everything. Reload that just doesn't for whatever reason resolve. And then each one of these represents a workflow that I threw at it. So, these were simple linting things that I wanted to have it go and fix for me. It worked out pretty good. It then goes and it creates a branch as you saw before. That's AI on GKE. This is I got a lot of things going on, but it will go and actually create a new branch based on whatever the summary is that it creates after it does the work. So, if I wanted to go and make a new one, like I said, I could open up that pane. There's also a little check box at the bottom left, which you can't see my face. My mug is over it. It's right here at the bottom left. There's another way that you can access it. Or if you have the pane open, then you can click on the cloud at the top and it'll say show background agent history. And then you'll see the diffs as well. So, it is in beta, but don't worry, it works. It'll just go and do all this stuff. But that's why I recommend having really strict linting in place and formatting. And then that also leads into the CI and CD when you actually push up the code. Like once once it goes to make a branch, you're going to want to have more stuff in place. And for me, what that looks like is its biome. It's having patterns in the codebase that it can reference. So having things that are pure, having things that are very simple and smaller files. So more functional rather than OO. And just that goes back to the architecture way of things like having things separate and very small. But I found them to be decent so far. I'm excited to see where they go. But now I want to talk about these loops. So what is this codebase? I mentioned before, right? It's the codebase for my community that I run. And it's essentially going to make it so that you have a Discord bot that uses all the context of certain message histories. So Discord can get crazy when you don't go on to it for a while. So, if I would come back to Discord and right now, it's a pretty simple setup where we have a bunch of different channels with different conversations in it. But if you were to come back after, let's say, 3 days, then you might be a little bit behind. You might want to get a summary. Okay. What if you came back after five days or a week or two weeks? That's going to be crazy, too. So for certain channels we'll have this app and it's not deployed right now but it's name is Val and so Val is the co-pilot for Vibe with AI the community and Val is able to get that message history and then it uses under the hood vector embeddings through vector search on GCP. So you can see that it's all set up in here. So you have vertex AI, you have cloud login, cloud build, cloud run admin, secret manager, all these different things are products within GCP and I just want to walk through how it works. So if I go into my vertex AI and then I go into vector search, you'll see that you have indexes and then you have index endpoints. So this is the way that it's set up in Google where you basically you know one click to create the index and the index would represent the essentially space for all the message history in a certain channel to go for it to have numerical representations. So then it can do semantic search or finding things that are the nearest neighbor based on the settings that you have and how many dimensions you have. But ultimately, think of it as a database that turns those messages into numbers, right? That's what a vector is. And then you have an index endpoint that aligns to it. So then we have this endpoint and then val will hit that endpoint. And Val has a few different commands once I push it live where let's go into moderator only. So if I do forward slash and then I go down you can see LLM and this is just a normal LLM call. This is for me to understand how the system works. Then you have collector status. So collector status is the message collector system. So it's going to currently be proact or as we continue not looking into the history of all the messages. Next I can do that where it's going and scraping all of them. We have other ones like the semantic search. So you can search for messages using semantic similarity because sometimes people are looking for different tools but they may have a typo or something. But where this gets interesting is the summary. So I can generate enhanced summaries of conversation history. This is kind of like meeting notes but for Discord, right? And then you can type in the number of days that you want. So it defaults to sav. So if you just did slummarize on a news channel because we're going to have Val, we're going to have another agent that is the news agent that can go and takes trusted sources. So not all of us have to crowdsource it. We just have a news one. The news it's going to be birdie. Birdie is going to be who gives us the news. Then birdie's going to be in this noisy channel. Birdie's going to be just posting, posting, posting. And then maybe you come back and you just want Val to give you the goods. After a while, Val will also be able to do forward slashtlddr on videos. So being able to see someone that posts a video, I don't have two hours to go watch this new talk. I just want the TLDDR through the lens of a developer. You can also do a channel summary. So that's the one that's upcoming where like you could literally just take a full message history and have it do that. We have different types of summaries. So we have general overview, technical focus, decisions and actions and key insights. What else do we have for Val? Then user getting information about the user. So I'm pretty excited about the Discord SDK and what you're able to do with it. And under the hood for the code, it's been a little bit tricky because I haven't used any of this stuff. And so this was a weekend project, but for the most part, it is now deployed. I need to do some testing today, but I'm really quite happy with the result. I think I started it on Friday and it's Sunday now. So that's why since you've been gone, that's where I've been gone. And the way that it is set up from an architectural perspective and what I'll be doing now with these background agents for like small changes is I have this AI done folder or sorry AI directory and I have a backlog a doing and a done. And so these are all things that I've already done and I was using a prompt from my AI SDLC. So, I'm going to be releasing that as well today. We've got a lot of different things that I'm releasing today, actually. So, if you want that, comment AISDLC and I will reply with the repository invite. But AISDLC was originally a CLI that I built last weekend. And what it did was it just chained these prompts together where you have a pitch, then you have a PRD prompt, then you have a PRD plus prompt, then you have an architecture prompt. And this is trained on Python and TypeScript. And when I say trained, I don't mean like I'm not actually doing training. This is a it's prompted, right? It's not trained. It's prompted for Python and TypeScript. What do I mean by that is if you're doing architectural decisions, these are all decisions and frameworks that or sorry patterns for Python and JavaScript. Now, these will work in other ones, but there's some other things in the testing patterns that will be obviously specific to either JavaScript or Python. Then we have, like I said, architecture system patterns. This is a self-healing document. So, one that will live with you throughout the cycle of creating features and then you have tasks. I usually use this. That's why I mentioned it. But for this, I was testing the cloud for capabilities of what does it look like without using every single one of those prompts. How good is it at getting the job done? And I wish I did use all of them. So I built all this in two days, but it probably would have taken one day had I used the AI SDLC because there were again architectural patterns in here that didn't make sense where it went and I wanted to see could it oneshot a lot of the architecture and it went and it did netify serverless functions because I wanted to use Astro and had I used the AIS DLC it probably would been like, "Hey, like you should have express for this one area because you're going to need it for XY Z reasons." I just said zed. Weird. So, let's look at the deployment architecture and some of the system architecture so you guys can understand how this build worked. So, let's see. I wish I could zoom that. Let's grab this graph graph and let's go to mermaid.live. Throw it in here. I want to talk about loops too. I'll do that in another video. But this is how it works right now where you have the client layer which is the client browser. So we have a landing page and just for the heck of it, let me just run this real quick. It is a turbo repo. and pmppm rundev. It's all dockerized as well. So let's see will this blend? Will it blend? Yeah. So we have the web server for or sorry the website and that's an Astro and I've got to add some stuff today. Looks like we have an invalid URI. Got to check that out. We have the website which I will be changing up some of the copy on it. And we have this Discord threading bit which I need to remove because I was going to expose eventually the Discord will be semi-public tied to ROS. So, look at a company like Midjourney and they grew to two million paid users and they also have a section for non-paid users on their Discord. So, the idea being that you could have some of the conversations in here, but yeah, we have the wall of love. Thank you to everyone in the community who said nice stuff. I apologize if I haven't been as active the last few days. I've just been trying to build this and get the AISDLC and a lot of other exciting stuff up. We have a basically a community repo now with a ton of projects in it. But this is an Astro and then I'm testing today where you will do the join button and the join button then get some more information about you and then if you want to join then you actually go straight into Stripe and then Stripe has it set up to then create the account right so create the account set the subscription set up the database so we have a database with Superbase let's go back to that mermaid. So, Discord client, then you have the Astro web cloud. It's or web app on cloud run. We have a clouduler because we have the summaries that are created on your behalf every week. So, I can change that as well, but that's so that people learn how the summary functionality works. And then we have the express API on cloud run. We have cloud storage for static assets. We have val the discord bot on cloud run. And then we have GCP secret manager. So that is just so you don't have a bunch of likeenv things laying around. It makes it easier for other devs to work on it once they figure out how to use GCP. Now over in Cloud Run right now, we're using 20 Flash to Grock through and basically be the LLM of choice. It's really cheap for going through all that stuff. We have the Vert.Ex AI text embedding 004. Then we have I don't know if that's right. I'm not sure which one it is. Maybe it is that. And then ve vertex AI vector search. And then over here you have the express API that then is dealing with the Discord API, Discord OOTH, Stripe API, Superbase Oath, Stripe web, web hooks. Yeah, all these. So this is a really fun build. I'm going to be testing it out today. And yeah, I just used some of the background agents really at the end to fix some of the like linting things. I haven't even merged it. So, I just I wanted to test. This is like my test for Cloud 4 and was really happy with the results of it. Now, I can see what else I guess I could go into news. But now, that's going to be it for today's video. If you found any of this helpful, make sure you like the video, make sure you subscribe to the channel, and you can check out our community. We're going to be launching all this. I'm really excited as well for the stuff that comes behind this, which is Oh, this is Medie. Yeah, Medi is a cheetah that is so fast and so good at fixing our code bases. So, this will be who notifies. This is like observability logging all that stuff for the discord. But eventually this will be dealing with log traces as they come out of any project and then have the agentic flow for fixing things. So that's cool. Also thinking about how you'd have different agents that are trained on different languages. So this would be more like you don't start with an agent, but if it's trained on Python docs and it can answer questions and break things down. think that Discord is actually really interesting to me as an LLM wrapper for a community where you could just have one of these that's sitting in there that is trained and you'd have a flow for the docs and the popular things like you'd put like Flask, Django, Fast API, Podantic, what else? SQL Alchemy like all the different things that are in like the typical Python stack. You just train it on an embedding pipeline for this and you can train it on code bases as well. And it's really easy in vertex now that I've gone and done the work over the weekend to figure it out where you could have a channel and then you'd have threading and you could actually speak to it in DMs and then yeah, you could have Python specific ones. Yeah. And then that's pretty cool. Now that's it for this video. I'll see you in the next one.