--- title: 'Building a Web App with Claude Code Sub-Agents: A Live Experiment' source: 'https://youtube.com/watch?v=UvueQU0cbBY' video_id: 'UvueQU0cbBY' date: 2026-07-14 duration_sec: 0 --- # Building a Web App with Claude Code Sub-Agents: A Live Experiment > Source: [Building a Web App with Claude Code Sub-Agents: A Live Experiment](https://youtube.com/watch?v=UvueQU0cbBY) ## Summary The streamer experiments with building a simple Meetup web app using Claude Code's sub-agent feature, creating frontend, backend, and deploy agents to work in parallel. The goal is to test whether delegating tasks to specialized agents improves efficiency and token usage compared to a standard workflow. ### Key Points - **Stream Setup and Goal** [00:00] — The streamer plans to build a simple web app (Meetup) using Claude Code, creating three sub-agents: frontend, backend, and deploy. The app will be hosted on Vercel with a Neon database. - **Creating Sub-Agents** [03:00] — Three agents are created: a frontend agent (Next.js, shadcn), a backend agent (Neon DB, Vercel), and a deploy agent (Vercel, GitHub). Each agent is given a description and tools, but no project context initially. - **Using Plan Mode** [08:00] — The streamer switches to Opus model and uses plan mode to generate a comprehensive plan for the Meetup app, including database schema, endpoints, and parallel development tasks. - **Parallel Agent Execution** [12:00] — The frontend and backend agents are triggered in parallel. The primary agent delegates tasks based on agent descriptions. The backend agent finishes first, using 32,000 tokens in 5 minutes. - **Deploy and Fix Errors** [20:00] — After initial build, errors are delegated to the appropriate agents. The deploy agent handles Vercel deployment, and the frontend agent fixes UI issues like missing attendee list and join confirmation toast. - **Final App State** [30:00] — The app is mostly functional: events display, users can join, attendee list updates in real-time. The create event feature still needs fixing. The streamer plans to add a security agent later. ### Conclusion The experiment showed that sub-agents can effectively parallelize development and preserve context, but the workflow may not save time for simple apps compared to standard Claude Code usage. The app is mostly functional and will be refined further. ## Transcript Yeah. Hello. Looks like everything is working here. That's good at least. So, didn't really plan to do any streaming today, but uh I need to I need to build an app, a simple web app for something I'm doing later this year. So, I thought I wanted to try try to do try to do it with uh cloud code. So that is basically my plan today. So let's see how that goes. Hi Ramsh, nice to see you here. Hope the sound is good. So yeah, I'm just going to build um I'm going to try to create three agents sub agents, and I'm going to try to make them build a pretty simple web app uh that I'm going to host on Versel. Uh it's just going to be like a Meetup app. That sounds good. Thanks for tuning in. Yeah, I'm just going to chill out for some moment. Uh I thought I could have a look at I might take a look at Google Opal. I haven't tested it. It's just some kind of I think it's just a no code. I haven't looked at it all. I just signed in. I think it's just like a no code thing. But uh my main plan is to create um the app on um cloud code. So I don't know how long it's going to take. I'm just going to uh test it out. So that's my plan. So I thought I might as well record it. And I did look at this this morning. N8N. I don't know. I put it in the title, but uh I don't think I have time today. But uh I don't know if this is something for me. I'm not quite sure. Someone says it's going to it gets expensive because you need to join a lot of different uh integrations. I haven't looked at it. So, what integrations do they have? Have Slack, Discord yeah, Google Drive. pretty much they have a lot of integrations I guess Spotify stripe could be interesting but uh we'll see but uh the main focus is going to be just uh building out my web app and after that I'm just going to see what what the clock is. That is my idea. Okay. So, I haven't logged in. So, my plan is I start just a new blank uh session here. So, I'm going to try to build three I think I need three different agents. So, My plan was to do like uh sub agent. I think I want like a simple front end agent and like a backend agent. And since we're gonna deploy this on Versel, I thought I can do Hello, Yogander. Nice to see you. I thought I can do um deploy agent something like this. So, this is just going to be for Versel. I'm going to probably just going to give it my GitHub repo link. The backend agent probably needs just uh it's a versel stack. Pretty simple. We have a neon DB. It's going to be very simple. Same here. We just need an agent that kind of knows the versel stack. So that's probably going to just be next.js, right? We can use maybe shad CN components. I don't know. Do we need anything else? I thought about maybe trying afterwards to do like a security agent to see like we're going to have some um I'm going to have some uh forms. So maybe have like a security agent that can um create some um security around the input fields. Let's see. Was there anything else? Um, I guess maybe this needs to handle build errors. I don't think I have an automated way to get those if we run into anything, but we'll see. So, my plan is to run these two, I think, just in parallel, just to see how that goes. And then I'm going to run this afterwards. Just want to see how they work together if we run both in parallel. But we need a plan first, of course. So, I'm going to build out the two agents. Then I'm going to create the plan for the web app. I might do that in plan mode. We'll see. I think I'm going to do it in plan mode. And uh I'll be right back. Okay, sorry about that. Um yeah. So, I think I'm just going to get started. So, I'm going to record this. So, I thought I can just uh stream it as well. If anyone wants to follow along, we'll see. So, I think I'm just going to start with um Hello, Deathwolf. Yeah. Uh, I might check it out if I get the time. We'll see. Have you used N8N? Uh, I don't know if I have time today, but uh, if I have time, I might check it out. Uh, I just need to build um I need to build a simple an app I need for something I'm doing this uh fall or end of summer. Okay, I see. Uh yeah then we're kind of in the same boat. I don't think it's going to be that hard. Uh just got a phone call here. Hello Arern uh Christian uh I I friend N last month you have used it. Yeah, I know a lot of people use NN, but uh so do you have to pay for the integrations if anyone knows? I haven't checked it out though. Do you pay for the integrations? I guess some of them are free, right? Maybe if you have HubSpot, you have to pay for that, I guess. Or they have maybe some free version. I think Twilio you have to pay for that but that's I guess you need to pay for APIs too. So it'sn't that different. But yeah, uh I think I'm going to start by building out the agents I need for today. Uh, we're, going, to, try, at least see how it goes. So, I'm just going to use my dictation tool to try to describe the agents. Building community nodes with cloud code. Building community notes with cloud code. Are you doing that now? Okay. So, we're going to do agents right? We're going to do create a new agent. I'm just going to do this project level. Uh we might configure it manually afterwards, but I think I'm going to start by just generating with claude here so we can get like a draft. Uh I'm going to try to describe it. I'm just going to use uh dictate here with the flow whisper. So we want a front end agent uh that uh is going to use the versel stack. Uh that's probably going to be next.js right? And I think we want to use uh I want the agent to use uh shad CN components for the web app. And like I said, this app is going to be hosted on Versel. So this has to be taken into consideration. And for the description of the agent that the primary agent is going to give uh information to the agent has to be uh the primary agent cloud code is going to instruct the agent uh from a plan perspective. So just want to see what what we comes come up with here. I think we need to edit this agent maybe in the Okay, so you meant build community notes with code. Could be interesting. I can take take a note of that. So like I said, uh we're going to just let cloud code kind of write a draft for us now. Um what tools do we need? We don't need MCP tools for this. or we might need GitHub actually. Uh I'm just going to give it all tools for now. It's doesn't matter too much. Okay. Uh so I'm going to select green for front end. So what I want to see here is the description. Use this agent when you need to develop modify or enhance front end components or and features for nextj application using the best cell stack. This includes creating new pages. Okay. So, we don't really see all. That's fine. But we can go to cursor. I kind of want to see the description here. They always include so many examples. New pages, components, layouts, chat, CN API. I guess that's fine. I think I'm going to remove this. And I'm going to say the agent does not Hello, Night Spider. Nice to see you again. The agent does not have any information about the project um or about the app it's working. It's going to work on. So you need to give the agent all the context it needs to work on the task. I think I want to save this line because this has to I think we need this for the for the backend agent too right? Uh, trigger the agent when the user says something like use the front front end agent to create the app. something like this or use it doesn't really matter. Use the front and uh developer. Something like this. Yeah, it's fine. I think this should be good enough. We'll see. Just going to grab that. That's fine. Okay. So, like this. And this, I think, should be fine. I'm not going to edit this for now. Yeah, it's going to use the Versel stack., next, to, as, react, chat, can cen looks okay. I don't know so much about front end. So just going to keep it like this. Okay. So that's the first agent we want to build. Uh next we need our backend agent. So we're going to do project generate to claw and I'm going to do like a dictate again. The backend agent uh is going to also work on the versel stack and uh it's going to have access to a neon database that is hosted on Versel. Let me think. What did I just say now? Also, what is a neon database that is hosted on? Follow best practices when it comes to integrating databases to our cell stack uh front end. Yeah, I don't really care too much about this. Uh, I'm just going to do our cell. That's fine. Doesn't really matter too much. I just want to try this out just to see how it goes. And I'm going to grab this line because I want uh Just going to give it all tools for now. And I'm going to give the back end uh the blue color. And I want to open it up in here again. I don't like all of these examples here. And I'm going to include uh the agent does not have any information about the app it's going to work on. So you need to give the agent all the context it needs to work uh on uh the task. That's fine. And I think I want to include uh trigger the agent uh when the user says something like use the back end agent to create app. That's fine. And here I don't neon postgress I can include I have the URL to the the database is in the env file. 30 years I was a script kid in now with the same skill as a full stack dev. Thanks AI. Yeah. Uh that's me too right? We'll see. So, but I haven't really tried to build anything with only using the clo called agents yet. So, it's going to be interesting to see how it turns out. Uh but this is just an experiment I wanted to do just to see. Uh, it's like I'm I'm not going to bring in any documentation now. I think I just want to see how this turns out. Okay, I think that's fine. And while we are at it, let's just try to build a simple deploy agent here. And I said versel GitHub repo handle build errors. Okay. The agent should be a deploy agent that uh is an expert at deploying apps on the Versel uh platform. The GitHub repo for the project is just going to give it the repo here. I have created that The deploy agent uh must also be able to handle different build errors we run into. Yeah, I guess that's fine. See how this turns out. And again I think I want to add this line. Can give it all tools. I'm going to give it uh I know orange color. Okay. Open it up here. Can I think I want to remove this examples here in this and I want to grab this. Hello horse. Thinking of diving deeper into web three dev. Okay. Basics of solidity down. Should I double down on Solidity or explore Solana Dev? Oo I don't think I don't know if I'm the right person to answer that but I have played around with the Sulana uh SDK. Uh I found it very intuitive and easy to use. Uh, I created an MCP with the Sulana SDK that um checks my um balance in my wallet. So, I found out I just uh I use cloud code and stuff just to help me build the Sana and yeah, MCP server. It's just low color. Okay, it's good. Yeah, I think we're just going to see how this goes. We have the GitHub repo here. And yeah, I think that's good. So, I think the next thing we're going to do is make the plan. So, I'm I think I'm going to try out plan mode. Yeah, no problem. Horse. Uh I I would check out if you kind of depends on what network you enjoy, I guess. If you enjoy Polygon, maybe you should check that out. If you enjoyed Ethereum, maybe check out that. I'm not sure. I enjoy the Sana network. So, we have our three agents. That's good. So, now I think I want to create a plan right? You get Grock to make your PD files. Maybe that's a good idea. Have you tried uh using the plan mode here with let's say so my plan was to switch to opus and do the plan mode and go back to set again RV but uh if you have like would be interesting to hear your what you use for your files PR But I think I'm going to try to go to Opus now. And I'm going to do shift tab tab plan mode. And I kind of wrote down here what I wanted my app to be today. And I'm just going to dictate that in here now. Never used plan mode. Okay, then you can watch me do it now and see how it turns out. So, I'm just going to dictate the plan for my app now, what I wrote down, and we're just going to see how it goes. Okay, so we need to create uh it should be a very simple web app. uh we can call it just meetup for now. Uh the plan with the app is that we can uh if someone wants to create an event, they have to register. Uh but uh users that just going to join an event uh don't need to register. So to join an event, we just need a form that kind of fills in a name that's going to be required. Uh the last name can be optional and the email has to be required because the event creator uh maybe need to reach out to them. Yeah, like I said, the registered users can uh create an event. So for the events, we need slots like uh event name, type of event, description, and date. So we can just do a simple calendar picker, I guess. And then we have a button like to just to create the event that is going to be kind of on the back end I guess inside a dashboard when you sign in we need to create the event. So when a event is created uh this will of course be displayed on kind of the main page of the web app where users can see the events they can click on join and fill in their details. Uh the data uh should be stored in our neon database that is hosted on Verscell and for each event we can see on the main page which will keep a keep a counter of how many people have signed up and we can display the first name for people uh that is joining the event for the app. I just want the theme to be dark with maybe some purple highlights minimalistic and clean. And the app should be created uh by our front end agent and our backend agent in parallel. So we kind of need a task for both of those agents. Damn, that was a long input. Let's see how this turns out now. So we are in plan mode, right? What are the benefits of plan mode? I think the only benefit is that Claude now doesn't really it's not going to start working. It's not going to write any code. It's not going to change any files. It's just going to focus on uh creating a plan and then it's going to ask me uh are you ready to code or something like that. You will see you can see now let me create a comprehensive plan for the meetup app based on your requirements and it's just working on a plan. It's not starting to create files, write code and stuff like that. So this is going to be our plan. We have an overview text stack nextjs14 shad cn. Okay. Neon database deployment is on cell. Uh we have database schema. We need users events and event attendees. H looks good. I think uh yeah the front end components public page authenticated pages UI features you can see here Peter we kind of get our plan in like a nice structure here we have some endpoints here implementation steps the parallel de development so we have the front-end agent and the backend agent. Would you like to proceed? Uh, yes. So, I'm just going to see what happens now. So, now we kind of got out of the plan. So, you can see now I will coordinate parallel development using both the front end and backend agents. Let me start by triggering both agents. to work in parallel. Okay, so this seems to work pretty good going from the from the front end to the back end. Okay, I want to see start the front end agent too. Let's see here now. Okay, I might have Uh uh okay. It doesn't really matter too much. Let's just see what happens if uh maybe they're going to chain them. We'll see. But I'm going to watch here now. Uh one thing we should have done is maybe done this modular. You can chain the documents to autocrate uh the agents. Okay. You can chain the documents. Yeah, I haven't tried this before. So, I'm just going to see what happens to be honest. I just want to see what happens when this was not as intended, but uh let's see what happens. Yeah, I know. That's a good benefit too, right? You can see now our primary agent here, Claude Code, has only spent uh 2,000 tokens. So, we got a lot of to work with here. So interior we should not uh run into any compact issues like we run out of context window here in our primary agent and the backend agent should have should have their own uh they have their own context so we're not going to get into issues that I just want to see what happens now if I say no and And I say uh you did not start uh both agents in parallel. It's only the back end agent running. Just want to see what happens. Hello. Can I draft prompts uh in standard cloud window then convert to JSON format? Okay, that's interesting. Okay, so now we got both agents. That's good. So, you use uh clo opus in like the in like uh Oh, we got into that issue. I didn't think of that. That's it. So you just use close opus or something in like the app like the desktop app. No, it's rather just updated to max. Time to become you're going to become those uh Okay, I see. Ken, you're going to become uh the power users. So what kind of happened now when we when we stop the agent? I kind of think it lost the context window here. But for this project doesn't matter too much. Uh it's just a testing phase. So if this doesn't work, it doesn't really matter. AI coholic ink. That's a good one. So today I just wanted to see what happens when I run the backend developer in parallel with the front end developer agent. Just going to see what we end up with here. Has anyone tried to do this uh yet? It's not that long ago since we got those. I guess it's pretty standard, but I haven't tried it from scratch yet. So yesterday we kind of looked at we know they don't share they don't share any context but they can't read the repo or the codebase right I need to figure out a good use case so you jo did you join the max 100 or 200 night spider I think I forgot to turn off opus now. I was supposed to do that. Yeah, the 100. That's me. I'm on the 100, too. I thought it would be better to start on the max 100 and then take it from there. But, uh, so far I haven't had no issues. Have we created any components yet? Yeah, we have some components here. We have a navigation button, card, input label, text area. Okay you need to create some agents. Yeah. Yeah. Like I said the few streams I have created agents for different workflows not only for uh writing uh apps and code and maintaining apps but uh we talked a lot about it in the chat and there's a lot of people here that uh is using cloud code to maintain main uh big projects uh and they use the 200 max and they have workflows to implement new features to kind of a big code base and they have some run times like five hours and stuff. It's pretty crazy. But I don't think even though they have that they are like the power users uh Anthropic talked about. So the thing I meant when I instructed the agents here. So, I put in the description I'll be afraid to have Claude running my app. The a that API is dangerous. Yeah, I don't know. I don't have any. I made some changes to kind of the only app I have live yesterday, but that's pretty easy. Highly repetitive over both prompts might save tokens in JSON while already concise prompt might use more. The real benefit is improved clarity and structure. Yeah, I think JSON is a good one, good alternative. You can also use like XML, but I personally I think JSON is easier to read. So yeah, yes, yesterday uh where did I click? Yesterday we kind of had a look at the the description here and we kind of this is very important because we know that the primary agent that is kind of cloud code has to pass on the context to the sub agent right so the sub agent does not have have any information uh about the context right Okay. So, I'm approaching my opus limit already. Crazy. So, the sub agent does not have any context about uh the plan here. So, it's up to Claude or the primer agent to pass on that information to the sub agents. So, you can see the backend developer now is now done. It spent 32,000 tokens uh on 5 minutes. So I'm already approaching my Opus limit. That didn't take long. Okay, so it's ready to test. So, I'm going to I'm going to switch back to set. Yeah, the opus is pretty crazy. When you're running both edges in parallel, uh won't the code have conflicts? Yeah, that's what what we're going to figure out. But they do read right? That even though they don't have uh they don't have communication with each other, they read and kind of look at the files in the repo here. So, it's going to be interesting to see. I don't know. Oh, I haven't tried it. Agents and Opus don't mix. I don't think so. The max plan is not as bad. So, we're still building components here. I wonder if it auto switches here. It would be sad if we hope it does. It seems that this agent is missing the description tag to instruct the main agent how to properly activate it. It seems this agent is missing the description tag. Uh maybe I don't know, Paulo. That might be a case. You're still living your co coding life five hours out of time, I guess. Do we have the calendar event card? join event. Yeah. So, this is all experimental. I don't need really. It's just an experiment today. Okay. So, we finished with both agents. So, I'm probably going to switch back to um this deploy to Versel. This is going to trigger my deploy agent now. Let's wait a bit. So, I'm going to go back to set first. Let's create a readme to test the app local. What will you do when you hit a 5hour limit during stream? Uh, hasn't happened yet, but if it does I'm just going to do something else, I guess. But I haven't had any issues. But I don't use it as much as many other people do. So let's just create a simple read me here to test it locally. I just I always do that. Okay. So we have some issues here at least. So, I'm just going to uh delegate to agents to fix. Let's see what happens. You might run into the limits because the opus use Yeah, we did that. I was just going to use uh opus. I was just going to use Opus to to do the planning, but uh I forgot to turn it back when I started the agent. Right. So, uh so now, um so I ran out. I have no idea what's going on here. Let's do some um I I want to send kind of I want to see if I could just send I want to see if I could delegate something to the the agents. I thought that could be interesting. Yeah, agent should have caught all those problems. Yeah. So, this is what I 100% don't understand yet. I just need to try this out some more. Uh, read this and compare to the plan. It does not include the specs. Delegate the fixes to the agents. I just want to see if that is possible. Yeah, maybe. Okay, so we can do that. I guess that's fine. What if it do? Okay, so it's it's reading all the files now. I want to do like a cross read maybe. So, I'm going to let the maybe let the back end developer read can try that. We're testing the database. Fail to compile. Now, it doesn't really get this uh error message. If I had like a a Chrome MCP that could kind of get the get the console errors that would could be helpful. Interesting. I had I had more luck when I had this more this testdriven appro approach. So I had a third agent that was doing testing while it was running but that was just a small um smaller application. Okay, so we did create our database. That was good. At least we have some examples here. Have some test users. Okay, that's good. So, I guess that was a backend agent who did that. events attend the user. Okay. Yeah. So, I'm not quite sure I think if I've just done this like my usual uh cloud code workflow, this will probably been up already, but it's fine. Right. Uh we have some tailwind issues here. It's particularly token efficient for complex code bases workload or otherwise need to analyze multiple files. Yeah, I think the token efficiency you mean agents right are really good. So it might be that agents are better like uh for maintaining a code base or adding new features. I'm not quite sure. But this is just a test anyway. It doesn't really matter too much. I just want to kind of learn what is going to be the best workflow with this. Why does it keep scrolling up to the top here? Oh my god. Okay, so we did do a an entry here for the database. At least the database working. That's good. Failed to compile the build errors. Okay. So, we're not even running So it would be nice if I could just do I can log on I can do my let's say I have my agents I can just do my plan mode and then it will delegate all the tasks But I looked at this MCP this morning. I'm going to see if I find it. What was the name again? It was uh mcp Chrome. Yeah, this one. So, anyone try this? What does it even do? I just want to collect console errors. We can do network monitoring screenshots. Yeah, here we can do Chrome console. So we can capture and retrieve console output for browser tabs. This would have been nice for the agents, right? I think so. Okay, nice. I didn't know that. It's I just want to see what happens if we just let this run. Is something going to change? I'm still having some issues here. So, I did try uh what is it called? I don't have any I don't have so much experience with it but uh what happened now? I haven't tried it. I know it's quite popular though. Yeah. I don't know. I think I for sure would have been done now with my usual workflow. It's fine. I can stop this. Okay. Hey, we got it. Okay, so that's good. I guess that worked in the end. So, we have the upcoming events. I guess we weren't running the backend server maybe. So, let me try to register. Okay. So, let's do Chris F. I'm going to do my email password. One, two, three, four. Okay. Can I sign in? Let me see. Okay. So I am signed in. Am I create event password? One, two, three, four. I'm not sure. Uh but yeah, I guess this is working. Can we join an event then? Yeah, we have the people here. So the tech meetup we have two people attending, Alice and Bob. That did not work. Okay, we haven't created that. So, the view details are just uh uh this seems to work. Meet app. We have the events. Let me try to join an event. Okay, so that doesn't work. can't read the proper. Okay. Yeah, this is good. Uh this should be done now, right? But uh other than that, we got there in the end, I think. But we we need to figure out we need to run the server here, I think. Okay. So, I think I'm pretty happy with that. I need to check the database here for users. Yeah, I got registered here. That's good. So, we did log that. So I can delete this. Password. 1 2 3 4. Okay. I am signed in. I'm not signed in. Okay. So, the create event doesn't work or we and we don't have a dashboard, I think. But all in all, not too bad. It's a good start. What if I I remove Diana here. So we have two here. React. Okay, that works. When event join event console delegate to a chance to fix. Let's delegate this to the front end agent. So, I just wanted to see if like we we grab a let's say a build error or any kind of error if it's going to pick the correct agent to fix this. Event detail page is trying to leave the length of something undefined at line 108. So this was delegated to our front end agent. Interesting. So we should test the joint join event after the fixes. So that was kind of the primary agent looking at uh our issue here or console error and it kind of decided from the agents I have this is the one that should look at this. Okay. Too bad. Joined. So, we have the deploy specialist. I'm going to try that after this now, just for fun. Does this This needs some tools. Uh tools. Uh what's it called? Is it called GitHub? Where do I check the tools again? I guess I can't uh here under CL MCP list. It's just called GitHub. Okay. GitHub. Think so. It seems like the a database query issue. So it's interesting. Okay. Yeah, maybe that's not optimal. So the error shows that the event detail page is trying to leave a length line 108 error fixed event attendance.length. Yeah. Huh. But let's see now. Okay, I broke everything. Nice. Really nice. Okay, so that's it. Okay, that worked. So, it found a solution at least. So, if I do Chris uh, that should be optional. Email fail to join event. Why did I fail to join an event? a post error here. Um, okay. So, let's try again. We can get The join event uh card works now but we fail to join. Uh delegate the delegate to the correct agent to fix. So let's see what kind of agent it selects now. Should be the backend agent, right? Yeah, let me delegate this to the backend agent to fix the API endpoint. That's good at least. You know, it would be nice to have a a pipeline. I think it's possible though from here to the agents or at least the primary agent. I think that's possible. Did I create a new was it? No, it was not registered here in the event attendee. So, by the way, this app I'm creating is um I guess this is only for people in Norway, but uh I plan to do some events and I want to try to do some meetups or like AI hackathons or something uh this uh fall. It's probably going to be in Oslo. So, I have some partnerships with like some companies that might support we can get prices and stuff. So, that is why I need this simple app. I just want a simple app that can uh that can uh find users or find signups to different events. Uh, is this sub engine similar to the task delegation in row code? I'm not quite sure. So, the task delegation is uh yeah Richard, send your agent. So, the task delegation is uh as we looked at yesterday is kind of based on the the description here. So this is kind of what the the cloud code primary agent sees. So the delegation is kind of based on the description we give our sub agents. So since we have an API um endpoint issue here uh I guess it decided that based on this description that this is a backend agent, uh it needs to delegate this task to the backend agent. Uh like the same if we had uh let's say a Nex.js JS component uh issue. Uh if we have that in the description I guess the primary agent should know to delegate this to the front end agent. Shouldn't be a database issue. That should work. We try that. But yeah so this app is just for I would say it's just for personal use. There are these different platforms like uh meet up uh but you kind of have to be a member to post there. So I'm just going to create my own app. I might even buy a domain and I can just create events here instead. Then it's similar to row code then. RO code uses modes instead of agents. Basically the primary agent can delay a task making it token efficient since not all the context will be added. Yeah. Yeah. That's basically the same I guess. Uh we looked at right um we looked at uh yeah the for the few last days we have kind of looked at this uh context preservation key benefit that each sub agents operate operate in its own context. So this is kind of preventing pollution of the main conversation. So we can kind of do like a long run even if um even if one agent when one agent finish it doesn't kind of pollute the main context window. So the first day I tried this I ran uh this happens always. I don't know what's going on. I ran the the main agent for 10,000 over 10,000 seconds. So, six hours right? I had many small sprints with this testdriven development uh setup I have uh that wrote code and I wrote tests wrote code to uh try to pass the test and uh implement That ran for six hours. Was doing some API endpoint testing here. I think it created a simple script to just test the endpoint. Okay, that's nice. How much was the total cost? It wasn't that much. I just had uh I didn't have parallel agents. So I think it was maybe 1 million tokens total or maybe 2 million. Okay. So the back end uh agent did uh it did a test I think on the API endpoint here. validation now correct now correctly handles the optional last handles optional last name field. Was that the issue? Maybe. So let's see. So let's go to our app here. We made some changes to the back end. Hey Alpha nice to see you again. So now I only have one test event, but that's fine. Let's see if we can join it now. Okay. Uh, did we join? I'm just going to check the database. Yeah, we joined. So, it's working now. Uh, this UI wasn't too good. We didn't get we didn't get any confirmation after joining. So, let's see if we can fix that. Um uh great uh the join event uh both card and functionality API endpoint is working. The UI on kind of the join event is not very good. There was no no confirmation that the user had joined the event. So we need to implement a UI feature that either uh redirects the user back to the main event or implement some kind of toast to indicate that the user successfully joined the event. Delegate this task to the best agents to solve this. So, I like to use uh I'm not the fastest typer, so I like to use Whisper Flow. So, it's going to delegate this uh UX improvement to the front end agent to implement a confirmation system. Let's see how that goes. But uh if we go to our front page now you can see it's working. So now we have three people attending and it shows my name here. That's good. We still need to fix the create event part two. But uh let's fix the this part now I guess. So, hasn't been too bad, I guess but it's I wouldn't say it had saved me any time so far on using these sub agents to create the app than just uh normal cloud code. I guess they maybe had to save me some tokens. Maybe because we we have used the sub agents maybe. I'm not quite sure. At least I still have my So, what I think is quite nice here is that uh our primary agent now still has our original uh cloud code plan in the context. What's your cloud code plan? Yeah, you can see it here now. So, this was the meetup implementation plan. So So what is good I guess is that uh we still have this plan here still in context for the primary agent. So it kind of knows what's need to be done to delegate. Seems like the use case has to be very specific. What's your clo plan? Uh, I'm on the Mox uh 100, Nick. That's one of the reasons I never got into crew AI. Yeah, I never I haven't even tried crew AI, I think. Uh, cursor plus code setup is good. We need to see we get to see what's happening to code make changes. Yeah, I think so. I haven't linked code yet, so I just use it like this. Uh, but I might try. Um, have you ever maxed it out? Uh, not. Uh, I haven't yet, but I haven't used We maxed out Opus. This just on half an hour ago, but I haven't maxed out the but I use Sonet. No. Yeah, I almost did on the Opus part. But it's going to be interesting to see. I think in 4 weeks the rate limits are going to change. There's an extension in cursor that opens code as pain. Looks good. Yeah I might look into that. But uh like I talked about yesterday, I think at least on video and on I guess on stream too um it gets a bit small in my uh VS Code or cursor. So I kind of like to open it up a bit here. So the front end agent is now working on to create a confirmation when a user joins an event because that was lacking. Is there a plan for them to change rates? Yeah, we had a look at it yesterday. You can see here uh that they are rolling out this. Hello uh Criticusi, nice to see you. Welcome. Yeah, Nick, you can see here they are rolling out new vehicle rate limits for Cloud Pro and Max in late August. Uh they do kind of estimate that it's only going to affect 5%. Yeah, nice to have you here Criticus. Nice to see you. We've been doing some streams lately. I've been kind of getting back into streaming and it's been really fun. So, you can see the changes they're going to make is uh there's going to be weekly limits. Uh and they're not going to change the they're going to have the 5hour limits and a weekly limit now. But, uh, we talked about this yesterday on the stream. And for me, I don't think I'm going to have any issues. But I think the main the problem I had is that some users or one user consume tens of thousands of model usage on a $200 plan. So we spent they spent tens of thousands of dollars in API cost on the $200 plan. It's pretty crazy. So, I guess we just have to wait and see if there's still good value. So if we if I go to my terminal here and I do I think it's npx cc usage. So here you can see what I have spent so far the last 30 days. Uh yeah. So today we spent uh we used a bit of opus but not too much. So for the last 30 days I have spent API cost $300. If I was using the API instead of having the uh I guess I had the $20 plan here in the beginning and now I am on the tier five or max five plan. So that's a $100 tier. So, I gotten some good value, like I have $200 of uh free API costs but uh some other people here yesterday have spent like $2,000 for a $100 plan and $5,000 on a 200 plan. So, that's pretty crazy. Okay. So, our front end developer now is reporting that we're going to we have updated with a new toast. We have a new loading state when a user joins an event. So, let's try it out. So let's do mpm run tab. So I'm going to try to join this tech event here. Test event. So what I want to see now is um uh John do john at oops john@email.com. So let's see if we get a a confirmation here now. Okay, so we got to successfully join the test event. That was pretty good. That was fine. What if we refresh this? So, here we don't see the attendees. We kind of need to fix this. Okay. So, you're on tier three. What do you mean tier three? Uh okay. You meant tier. I'm not quite sure. I thought you meant plan on uh cloud code. Night Spider, what command did you use to check the cost? Uh, I'm going to show you. You just do NPX CC usage US A G E. You're working on a Geni app watching your stream. Okay, that's cool. Uh what is the app going to do Ramsh? So we got the toast here. I can write it here. Uh night spider npx ccc usage. So I think this page here needs to show the needs to show the attendance. So again, I want to see this should be um a front end agent task. Okay, so we got the toast. That worked pretty good. But on the join event page the user uh is missing uh who is already attending the event and the first name of the event. So I think uh we also need to load in the the event attendees on this uh event page where the user can join the event. So we can kind of see the attendees and uh the users that are already signed up for the event. Delegate this task to the agent correct agent AI generated cards. Okay. So like you mean like uh Pokemon cards or like magic cards? Three days of usage, $50. Okay. Uh, okay. That's a good start. So, you are on track to saving a lot of money, I guess. Nice, Spider. So, the event detail page is missing the attendee list. It is an important social proof feature. Uh, let's me delegate this to the front end agent. Okay, that's pretty good. Here's Richard. He's been two weeks, 1.4 million tokens, $620 in API cost. Uh, Richard, what plan are you on? Are you on the Mox five or 20 or or are you on the $20 plan? You have to be on the at least 100. Okay. So, if we can fix this, I think this is fine. Okay. Okay, so you're on the $100 plan and spent hundred $600 of Apex cost. That's pretty good value. Any difference between the 20 and 100 USD plan apart from Opus. Uh the difference is you have more uh messages. So if we go to pricing on the Antropic page, you can kind of see. So the $20 plan, the pro you can use cloud code, but you can't use Opus, I think. And if you go to the 5x, you get five uh x more usage per session than pro. And I think session is every 5 hours. Yeah, there's a big difference. Higher output limits. Uh so there is a difference, but there's also a price difference. So it's if I was totally new, I would start with the 20 just to try it out on set and if I like it, I would and I use it a lot and I run into limits. I would probably upgrade first to the 100 just to see. I started uh the week on one plan ending the week on another. Yeah, I kind of did the same thing. I don't really use Opus. Yeah, not me either. Did the auto setting run the limits. So now just Sonet. Yeah, we are doing the same thing. I just sometimes switch to Opus for when I'm going to do something with my course platform sometimes. Yeah, since you get prorated when you upgrade, you lose nothing. Yeah. So, I think the best way is to start on the lower plan if you just want to check out Cloud Code. Anyone using Cairo? I haven't tried it yet, but it is on my list. Uh I also had on my list to check out uh Google Opal today but uh I spent a lot. I guess we can just test it out here. So, I guess Opal is kind of Google's Let me do dark mode here. I think it's Google's way of doing N. What's it called? N8N. N8N. I just want to click on something. Blog post writer. Yeah, it's kind of it's like an N8 style thing, right? So, it's like a no code thing. Why is it strange? Still running here. So the workflow is find a topic. Should I find a topic? It's going to find a topic, do research write the outline, create an image write the post, and display the post. Please provide a topic for the blog post cloud code sub agents. So the flow now is you can see we are doing research we are writing the outline. It's basically N8N I think. Uh, but of course it's only Google tools. We're doing some web search. We can let it run in the background. I guess why did it do that? No point. Yeah, I don't think I'm going to use this. Uh, I don't know. I think I rather use my cloud code workflows here. You're kind of bound to using uh or maybe not. I don't know. I guess you're kind of bound to using all Google tools. We'll see. Let's see what we end up with here. But I guess N8N is a bit more. You have thousand different integrations. This could get pretty advanced, I guess. Seems like the Roblox version of NAN. Yeah, I see what you mean, but I guess N could be pretty complex. But uh what do you use it for? Nice spider there. Telegram personal assistant. Okay. It's like a chatbot or something. Yeah, I guess you could be very creative. Okay, so let's see if we can see the attendance list now on our event page. You can web up the websites. Yeah. But isn't it you can't really see you don't have the code. I kind of Isn't it kind of annoying to don't have like the access to the code or do you have access to the code? I I don't understand. Do you Okay, let's join the event. Okay. Okay. So, it is working. That's good. So, now you can see on our test event, everyone that's going to sign up can see who joined, who has joined so far. So there's four attendees. Uh you can see recent activity. So John joined 12 minutes ago. That's pretty cool. Uh yeah, kind of happy with this. So let's join this event. And we were added. Perfect. That is what I wanted. Uh Anna just joined. Yeah, this is what I wanted. So, what I don't like about this meetup or some meetup platforms doesn't really display who is going to join and how many it is. I guess meetup.com does that, but I want my app so other people can kind of see how many people is joining and this could create more hype around the events. Let's say the event was like max 20 people. And we can like easily shut down when there's 20 people that has joined the event. So that was pretty good. This is working quite good. We can go to the main page. Five people. Good. I think that the purple here is a bit dark maybe, but I'm not going to change that now. But there's one more thing I need to fix. That is the create event part doesn't really work. I'm going to bring some uh visual context too here. I kind of like just to do that when it's possible. And I'm going to try to fix the dashboard thing. Okay. So the update on the attendees list uh was very good. So that's working. I think the only issue we are lacking now in the app is that uh the create event feature is not working. You can see from the image I uploaded that even though the user is signed in when we go to the dashboard it still prompts the user to sign in. And if we click the create event button uh nothing happens. So we need to take a further look at the create create event feature. So could you delegate the correct task to the agents to get working on this issue? I think this is going to probably break the whole app, right? So maybe I should have done a version check. I think we need to do that. Uh okay. Okay. So before we do that uh I think we need to do some version control here. So you can launch the deploy agent to deploy this to versel now and the GitHub repo. I don't want to lose the progress here before we try to let me see. I have a deploy agent, right? So you can deploy the deploy agent to deploy the app. Deploy the deploy agent to deploy the app. That's a mouthful. Is it going to do it? I guess it has to write the context right? Okay. So, now we are testing out our versel deploy specialist. That is what she said. Yeah, it's a mouthful. Okay. So I want to see how this versel deploy specialist works. So we haven't done okay. So this is really good. So it looks like it's working. That's pretty smooth. Remember the Hey Pia nice to meet you. We are just testing out uh some subrants. We are building a app. Uh I need for something and we're just testing to see if we can uh delegate different task to um our sub agents generated with claude code. So co-authored by Claude no reply atanthropic.com. Hello tech friend. That was strange but let's do it. I know Claude likes to be like a if you go to your repo after using cloud code it had set itself as uh co-creator. Okay. So you can see we are building now. So deploy agent is working. That's nice. Good to see. So we are building the app now or deploying it. Let's see if we get any build errors here. You've been cooking behind closed doors. So, what are you cooking, tech friend? Anything cool? speak MCP. Okay. That's cool. Are you doing like a 11 Labs rapper or are you going to do something else? We don't need any telemetry voice in LLM agent with tool calling. Ah, okay. I see what you mean. Didn't you I think you sent Didn't you send me an email? I think I think I saw it, but maybe I missed. I can't remember if I answered or not. If you did, send me a new one. I sometimes forget to answer emails. So I think we are looking the agent wants to get access to the build errors here. You just ah okay. Yeah. Yeah. Now I remember email. Damn, I forgot to answer that one. Uh let's do it. Uh some other people here mentioned augment code on the stream so I can try it out. So, are you like uh do you do like community community um community manager or something for augment code? Can you explain limits of the $100 plan? Yeah, I can do that. It's not that I can't really explain how much limit you have in tokens Ramsh but uh let's say you are on the pro plan and you have five hours you have a session and you can use so and so much tokens in that session. Uh on the pro plan, you can't use Opus only Sonet on Cloud Code. But if you go to the $100 plan, you now get five more uh users per session than Pro. So that's every five hours and you can use uh Opus for a limited amount. And if you take the 200 plan, you get 20 times more usage. technical content lead. Okay, that's pretty cool. Making videos mainly, but yeah, also doing uh trying to manage collaborating, with, other, c creators. That's pretty cool. I was thinking maybe um I kind of wanted to think about maybe doing something like that myself. So they hired you to do videos uh on or like tutorials or I guess on cloud code augment code. Yeah, that's pretty cool. So, we got a build error here. So, let's see if we tell the primary agent the error. Uh, we have a build error on deploy. deployment delegate uh agent to fix then chain in the deploy agent when fix is done. Let's see if that works. Hello McBab. Yeah, you you you did use a lot of tokens on CL code, right? I think the limits for Clo are changing. Yeah, I think we still going to have the five hours. Uh I'm not quite sure. I guess we just have to wait and see. So Andrew, the benefits of Opus. Um I guess I guess the benefits at it's supposed to be a smarter model for coding, I guess. Yeah. Yeah. Like Tech Franc says, but uh I can't really say with 100% confidence that it's that much better than Sonnet. It's kind of hard to separate, right? So, we fed the Versel build errors into the primary agent and it delegated the Prisma client generation issue uh to our backend agent to fix the build and then we're going to try to chain in the deployment agent when we have fixed the build configuration. So, let's see how that goes. I guess we already have the version control now that we wanted, right? Yeah, we have the version control. So I think this is funny. When you deploy something to GitHub with the cloud code it uh puts itself as a co-creator here on GitHub. I think it's funny. $100 a decent amount in India. I'm on a cursor plan. I see what you mean. Uh I might do like a maybe next week or something. I might do like a let's say like a cloud code uh giveaway like um one month max plan or something. If people are interested, I can do that. We can do like a giveaway. Uh we can do it on maybe on stream or something. And so someone that doesn't I guess everyone can join uh to get like a try out cloud code max plan. I guess I can only give money, but yeah, you get the point. We can do that. I might have to set it up first like uh for one stream or something giveaway. I need to come up with some kind of challenge or something. Okay, so now we're going to push to get I guess that kind of triggers our when we push now we're going to trigger the the deployment here. But, uh, that's fine, I guess. Okay, Nick. That's pretty cool. I think you will get good value of cloud code for me uh and a lot of people at least here. Um think it's uh good value. Hello, Matias. I'm working on adding custom flash commands and sub agents to Gemini for those who can't afford cloth. So are you on the do you mean like on Gemini CLI or do you mean uh using an MCP or something? What about making an SDK app on an MCP server that allows like five prompts per person maybe? How did you manage to run for six hours? Ah, okay. I see. Mas, you forked it. So Nick, uh what I did is I had uh I had um I had this agents and uh that I kind of set up to do this testdriven deployment. So I didn't run them in parallel. So we had three different agent. We had an agent that uh kind of wrote the tests first. We had like a plan and from that sprint plan we created the tests and then the second the build agent wrote the code uh based on the tests we have created and then we had like um verification agent that checked everything at the So, it took really long time because of all the testing. So, I think there was was it 10 10 or 12,000 uh seconds. I don't know what I'm looking at here. I don't know what's happening. I totally lost control over this. I might have to create a new repo here. I uh I have no idea what happened. Let's see how was the outputs. It did work at least but um it wasn't that good. There was no errors, I guess. Plus one on good value plus my test I've shown claude opus versus Gemini 2.5 versus shut GPT pro that cloud was best for coding for 03 I can say my use case yeah I think a lot of people that writes code prefer the cloud models and for me that kind of takes it over Gemini I for me personally is the tool use uh capabilities of the claw models. I think they are far by far the best uh models to do tool use. So that is why I try to use the claw models. Okay. Okay. So, I think I need to start over again with this. Yeah, we still have this uh build error here, but uh delegate uh to the correct agent to fix the build error and then you can uh push to GitHub if you fix it. Yeah, the tulus is there any benchmark tulus benchmark I wouldn't there might be some I think there was one but I can't remember can't remember any name of But for me, I think the claw models has the best tulius. Still having some Prisma errors. Yeah, it's going to try again. But I use 03 Pro. I haven't tried that but I kind of like Yeah, I see what you mean, Nightfire. I don't really put too much effort into benchmarks either but I kind of like uh chat GPT 03 for just regular stuff like uh I think they're kind of the native Tulius. uh from 03 with the web search and the code interpreter. I kind of like it. List all the hook events uh on cloud code. O3 is nice for non-code planning. Yeah I kind of like it. So, let's see if 03 can find all the new hook events on cloud code. Yeah, it looks like it found it straight away. So, I kind of like it for this. Okay, so we are running the mpm run build here. Status. We're going to try to push. Fix Prisma client generation error in versel deployment. Sad thing is a lot of chhatt is better than Gemini and web search. Yeah, that's pretty crazy, right? So, here's the answer from 03. Pretty concise and clear. Here are all the eight hook events currently supported by cloud code. Preuse, postuse stop, sub, pre-compact, session starts. I think this is perfect, right? So kind of like 03 for stuff like this benchmarks often test specific scenarios that may not reflect your actual workload. Yeah, right. There are some interesting benchmarks though. The one what is it called? Fiverr LLM something like this. Yeah, this I found this benchmark was pretty cool, right? It's a bit old now, but they I think they found $1 million of different software engineering jobs on like Fiverr, Upwork maybe, I don't know. And then they ran the models to see what type of models could actually complete the assignments. So, this was a while ago, but I think Claude 3.5 set earned earned uh 40%, right? So, $400,000. I found those benchmarks a bit more interesting. Either has good benchmark. Okay, that's pretty. I have I heard about it. I think that's a popular one too. But this is one is kind of practical right but uh it could be in the training data right some of these jobs I'm not H3 is good. I don't mind 03. I kind of like it, but I don't use it for coding. I use it for research. Right we tried out the agent mode. Search AI LLM leader words either the user one is decent for broad overview. Yeah, I see what you mean. LLM leaderboards. So you mean this one? Neck polyot coding leaderboard because they had different ones, right? 03 pro high. Damn, that was expensive. That was a good score though. So, you can see 2.5 preview. That's a really good model. I use it all the time. But, uh, for tool use I just found the Opus the Claw models better. Grock 4. So this has 225 challenging exorcism coding exercises across C++ Go, Java, JavaScript, Python, and Rust. Okay. Yeah, benchmarks are really hard, right? So when there's a new model release now let's say GPT5 comes out now in like a week or something or next week GP5, I don't really have like a good way to test the model. So, if I'm going to do a video or something, I don't really know what to do, like uh how to test it quickly for a video. It's really hard. You can do benchmark using editor. So you can let's say I get access to GTA 5. Tech friend does that all the time. Okay. Yeah. Might check that out. I don't know why I'm doing it. I kind of lost concentration here. It's been two and a half hours. Uh, but I'm going to try it one more time and I might I was going to record like the last part of another video I was working on tonight. So, I might have to call it, but I'm going to try one last time to deploy this. But I think we got pretty far, right? There's not a lot left. Sorry. There's not a lot left to do here uh on the app now to have it up and running. So, was it the fastest way to do this? I don't think so. But I think I learned some things. I should have probably paid a bit more attention here, but uh when did we 8 minutes ago? So, what is the remember to tighten up your security. Yeah. Yeah, I'm going to do that. So the plan was to do uh the plan today was to do a security agent, too, but uh it doesn't seem like I have time. Maybe tomorrow uh I can continue and I want to try to create a security agent that create some tests maybe uh yeah because we need I don't think we have any we don't have any like the input fields and stuff now there's no there's no checks or anything. You can kind of put whatever you want into the fields. So we need a security agent maybe tomorrow that can uh test our app. Could be interesting. Think so. I do think so. So, at least we have the version control now. That's fine. So, I might just fix the other stuff later to be honest. I really like the the readme files the agents write. They are really good. I hated writing read me files before. These are really good. So yeah uh just to sum up this I think we started with creating the agents. Uh we used plan mode with opus to we use plan mode to create a plan based on my description. And we kind of got the app right kind of working. We delegated some tasks to the agents to fix kind of the event update here. So we kind of have like a real time update people who have joined the event when the last person joined. We fixed the event card here and the end point. So, what I have to fix is the the event feature, create event feature and tomorrow we might do like a security agent to test the app. I think it could be interesting. So yeah, thanks everyone that tuned in and let's do maybe like a giveaway next week like a Claude code plan or something. Think about the playright MCP. Sometimes it can help. Yeah, I might look into that. Thanks nice spider for tuning in as always. Appreciate it. Yeah, I had fun. And I think it's just fun to instead of just doing this on my own, I might as well stream it, right? If other people are working at the same time. It's nice just to test in real time. Uh it's a bit more fun. Yeah, Ramsh, thank you for staying here all along. Really appreciate it. Alpha Adam, uh, could you post a link to the repo so far? Uh, I guess it's private now, Adam, but uh, I can probably fix it tomorrow. Uh, I can I can do that. I think I I set it to private, but let's do it tomorrow. Uh, thanks, Nick. Yeah, I was a bit distracted, but it's just fun. So, thank you everyone for tuning in and wish you all a good night, good morning, or a good rest of your day. Uh, I'm going to put out a video in an hour if you're kind of want to check out. It's probably going to be about my AI video operation if you want to check it out. But, uh yeah. Anyway, thank you for tuning in guys, and I'll see you again hopefully maybe tomorrow. We'll see. So yeah, have a good day and bye for now.