Claude Plans, Gemini Designs: One Workflow for Beautiful Frontends (LIVE)

All right, today we're going to be doing some experimentation. I've been testing out the new Gemini 3.5 Flash over the last week, and it builds some pretty crazy good-looking frontends here. But, uh, one of the things that it didn't do that well, so like this is an example of a a front end that Gemini 3.5 Flash built for me in the new anti-gravity. Uh but one of the things that it didn't do that well is uh it kind of hallucinated a lot of information on the on the page itself. Right? So it's like Gemini is super good at creating beautiful looking frontends, but then when it comes down to like planning the actual content and any kind of like integrations with the front end, it really falls flat on its face. Like the performance of the model just goes down significantly. And so everything that we're doing in this live stream today revolves around how do we combine multiple different models together in one workflow so that we can, you know, extract the parts of the model that perform best and then kind of diminish the the downsides of the different models. And so I have a diagram here that describes what we're going to be doing. We're going to be combining Opus with Gemini 3.5 Flash to create frontends that look good and actually have the right information. And so one of the things like with Gemini 3.5 Flash, I tested building a landing page that just like explains Archon, my tool, like how to run it, how to set it up, and it just like got so so much of that just completely wrong, but it looked good, right? So what if we have Opus decide here's the information to put on the front end and then we have Gemini 3.5 flash build the UI and then any kind of integrations that we need with our application uh like authentication for example or hooking into any kind of like backend then we'll use Opus again. So we're mixing providers in this larger workflow and of course I'm going to be using Archon to orchestrate everything here. So this entire flow that we're seeing here, we're going to build it as an Archon workflow live and we're going to see the kinds of results that we get. And so I have an application in mind that I'm going to build completely end to end. I have the prompt already laid out for it just so that we don't have to watch paint dry as I craft the prompt for the actual application that we'll build. But we're going to build a full front end using Opus for planning Gemini to build the UI. And then as far as integrations go, we're going to be setting up authentication with Clerk. So, I thought that'd be like a really good example of something kind of complicated that we wouldn't really want to lean on Gemini to set up, but like we need the front end first, then we'll add the authentication on top. So, just trying to make it so that it's a very non-trivial task that we have this workflow handle for us. And then, uh, even throwing in sonnet a little bit here for some validation as well, just to take it even further. So, really, we're using three different large language models in this workflow here. um using each one for I mean basically Sonic because it's cheaper and faster. Opus because it's the best at reasoning and then uh Gemini because of the UI. So that's what we got going on today. And uh before we like really get into it, I'll pop over to the chat here and uh see if anyone's been saying anything. Uh not yet. Cool. But yeah, for those of you who are here, glad to have you here. Uh I kind of messed up. So, yeah, there there's 40 of you here right now, which I appreciate, but I messed up and I scheduled the stream for 900 p.m. Central time instead of 9:00 a.m. Um, so definitely there's like a slower trickle of people coming in now. U Okay, the numbers are going up. So, cool. That [laughter] That's my bad, guys, that I scheduled it for uh the wrong time. So, I schedule it like from my OBS, from my uh broadcasting software, and so like it's kind of harder to see. I don't know, like scheduling a live stream from YouTube is easier, but then like OBS is just where I run the stream, so I just like set it up there as well. So, yeah, that's a mistake I'll never make again. That's funny, though. But anyway, let's let's get into it here. So, uh, yeah, good to see you, Sydney and Eric. Hey, guys. Now, now the chat messages coming. I think it might just be because of the delay of the stream, but yeah, good to see you guys. Um, and actually I'm curious like as I get into building the workflow here, have you guys explored Gemini 3.5 Flash? I'm curious if you guys have thoughts on it already or specifically if you've used it in the new anti-gravity cuz this is where I was first testing out building the UI like um the one that I was showing right here. And like I'm pretty impressed with it overall. Like I said, like the UI is good, information is hallucinated a lot. Um, so I'm curious if that's your guys's experience as well or um maybe you were able to like make it so that it was able to handle everything end to end. I'm sure I could have handled things better like more of a planning workflow so it's not just hallucinating information but it's kind of crazy just like building this single landing page. There's not that much here and there was like 10 things that were just like dead wrong where I totally didn't understand the tool even though I gave it access to the repo to like dive super deep into it and everything. Um, yeah. So, any that's why I want to use Opus here. So, I'm going to uh jump over to my IDE because I have things prepped already. So, I did quite a bit of work to prepare for the live stream here. Like, we're still going to be building the workflow from scratch, but I did a dry run to make sure everything works, especially with Clerk cuz uh we're we're picking Clerk to build like full authentication. There's quite a bit that goes into that. And then I did create the spec for the site already. And so as far as like what we're going to actually build, we're going to start by building the archon workflow. And so like this diagram is going to turn into that YAML file for the archon workflow. And then we're going to run that against a spec that I already have designed here. So actually I'll I'll pop that up here just so you can see it quick. I'm going to build a a benchmarking dashboard. So a tool for archon that allows us to run different workflows and basically like compare the outputs, right? So like we have a workflow that we want to test with different models and um then we can like run multiple versions of that of that workflow with different models and then have an evaluator that sits on top to judge the output of each one of them. Um, so this is kind of related to some of the other live streams I was doing where I was uh benchmarking like how good is Opus versus Kimmy and like what parts of the workflow can we get away with using Kimmy so it's faster and cheaper versus when do we really need Opus for that extra reasoning power. And in that last live stream I did where I I did this benchmarking, it was actually very interesting. The results showed that really the only step in a workflow that you actually need opus or like the best of the best is in the planning. So if you have opus for planning and Kimmy for everything else like exploration and implementation and validation, you get almost as good a results as if you use opus for everything. Like that kind of benchmarking is really interesting to me. Like trying to figure out how could we make things really token efficient, especially because of how expensive the best models are getting now. [clears throat] And so I thought this would be a cool application to build and then we'll slap authentication on top of it with Clerk as well. And so there's quite a bit that goes into this with Clerk because we're going to be using they they've shipped a new skill that made me like extra interested in incorporating it here. So we have a skill for Clerk that basically tells Claude Code how to set up all of the authentication for us. So, we'll be using Opus to handle all that after we build the front end uh with uh Gemini 3.5 Flash. And then as far as um how we're going to be using Gemini 3.5 Flash, we're going to be using Open Router. And so I uh did a lot of testing with just going right to the Gemini CLI directly or like the Gemini API directly. and I kept [clears throat] getting rate limited. Like my API key was stuck on the free tier even though I upgraded my my Google account so that I'm on like the second tier or whatever. It was really annoying. So I was doing a lot of prep here. I'm glad I did the prep beforehand and it like never worked. So using um open router to get access to uh 3.5 flash through pi is the way to go. And then we have the PI integration in archon. So not only are we mixing different models, but we're actually mixing providers. We're going to be using cloud code in the archon workflow for pretty much everything except for when we're using pi for building the actual UI. Pretty cool. Um, so yeah, as far as building the workflow goes, let me jump back over. Um, sorry, that's the wrong diagram. I have the build prompt right here. So, I'm going to send this off to my second brain to build the workflow. And then I'll kind of explain parts of this as it's building. So I sort of want to do both in parallel just because might take a little bit for cloud code to build this workflow for me. Um so yeah, we got like the eight steps here. Put [clears throat] a lot of work in to make and I I like iterated on this quite a bit. Um so let me actually copy this and then before I send it off, I'll jump back over to the chat because I see you guys are being a bit more uh talkative now. All right. So, let's see. Hey, Jason. Good to see you. Thanks for doing this. You're very welcome. Yeah, I'm excited for the stream here. Uh Sean said 3.5 and anti-gravity is not too bad. Not swole, but not bad. Yeah. Cool. Uh hey, Victor, you're very welcome. Appreciate it. Um let's see. any recommendation for courses on harness engineering and orchestration and um Ben said you don't need a course just learn what there there are many videos out there for it and um I would say yeah that's true like personally I'm doing a lot of content on harness engineering on my YouTube channel right now and I mean really like even this live stream is teaching you how to build harnesses I mean more specifically with Archon which is my open- source harness builder this really is the star of the show here that allows us to mix provider so easily with what we're about to dive into here. Uh but yeah, I've got some other content I put out on my channel um on harness engineering. The past couple of videos have been about creating harnesses. I don't know of like a complete course on building harnesses. Um I mean honestly the closest thing that I know is the agent coding course that I have in the Dynamis uh community but also uh it's not like ultra focused on harness engineering though that is going to be a huge focus when I do a revamp this summer of the agent coding course in uh Dynamus. So I'm excited for that too. Why [clears throat] am I sticking with Kimmy given the issues that I have had? So, uh, yeah, what you're referring to here, um, I talked about this in a couple of live streams recently, is that, uh, Kimmy the API gets stuck sometimes where it'll just like hang there forever as it's waiting for or like my Archon workflow will just be waiting for the Kimmy API to send a response and so it's not like there's an error so it can retry. It just gets totally stuck. And I actually have been experiencing that less. their API has been more reliable for me recently and uh even before it was like really really reliable, I wanted to keep using Kimmy just because like it is the best as far as like a super cheap model. Like I I it definitely got better results compared to using uh Miniax M2.7. I still want to do some more testing with some of the newer Quen models. So, like I'm not like finally set on Kimmy or anything, but it's just like a really solid cheap model and I can use my Kimmy for code subscription that I have as well um whenever I'm running Kimmy through Pi like with Archon workflows or just using Pi in the terminal. Um what's the [clears throat] reason you pick Gemini 3.5 Flash? Is it fast and cheap? So, the reason here is because it's super good at building frontends. That's that's literally the only reason. I mean, it is fast and cheap. Uh I think it's like I don't know the exact pricing here. Actually, let's let's jump over to Open Router so I can show you the exact pricing. So, if we look at Gemini 3.5 Flash, um it it is $150 for every 1 million tokens and $9 for every 1 million output tokens. And so that's not like insanely cheap. Like Kimmy is still cheaper than Gemini 3.5 flash. But if we compare it to let's say Opus for example, the difference is insane. It's um yeah, like $1.50 for input and then it's $5 and $25 for output tokens. That's not even the fast mode that they have. Uh Anthropic also offers that. So uh yeah, the pricing on that is like actually crazy. If you if you go to Opus 4.8 it fast. It's uh 10. Yeah. So, it's like twice the price. Um Yeah. So, anyway, very very expensive. But like Gemini 3.5 Flash will build better looking frontends than Opus. Like what it built right here. I mean, obviously there's probably a million ways you can tear apart this UI that Gemini 3.5 Flash built. But like this looks pretty dang good. I got to say I didn't iterate on this at all. it oneshotted this entire thing after just like digging through my archon repository. And the most important thing about this entire site is uh and may maybe I'm just being naive, but like this doesn't look large language model generated. Like there's a lot of websites out there. Um I mean I guess like some things like specifically LLMs love to have this like rounded box like smaller box um at the top of different sections of a site. So, like there are some things that are still kind of like, oh yeah, an LLM made this. But like overall, it's not nearly as bad as usually what you see. Like generally when a large language model generates a site, it looks like they're just using like default, you know, Tailwind CSS configuration or like material UI configuration and uh the you always have like the purple colors. I don't know, there's just always all those telltale signs that it's like everything was just vibe coded here. And I just don't feel that with Gemini 3.5 Flash. Like it feels like someone put a lot of thought into how to represent all the information. And that's why I'm really excited to build this workflow here where Opus plans what we are going to say on the site. And then Gemini 3.5 Flash takes the what and turns it into the how. Like how are we going to present everything. So like Opus would maybe say like, okay, we want to explain what an archon workflow can look like. like we have the plan, the self-healing build, we execute the tests, um, whatever, and then like Gemini would take that and then design this widget with that information. So, Gemini by itself can like make this look good, but like maybe some of the specific details here are like actually wrong, which I did encounter that when I was building this. And so, that's why we won't use it for everything, but just the UI. Hope that makes sense. Archon spicy mango version. Yeah. Um, you know what, Sean? You'd be really pleased to know that I actually ate some um chili lime spicy mangoes a couple days ago, and it was so good. Uh, let me see if I could actually Chili lime spicy mango. It Oh, man. They were so good. Yeah, this is This is uh Oh, no. That's not the bag. I got I got them at Target. Oh, yeah. Here it is. Here it is. These I had these a couple days ago, and it was it was so good. So, um [laughter] I would highly recommend Fruit. They're not sponsoring this video, I promise. Fruit Riot Spicy Mangoes that Yeah, I down the whole bag over the course of uh two days. [laughter] Highly recommend. All right. Anyway, back to our regularly scheduled program. [laughter] There we go. Oh, man. Join join Dynamus this week. Highly recommended. I appreciate it. And welcome, man, to the community. Yeah, we got a lot of exciting things going on in Dynamus right now. All right. Um, the community workshop in Dynamus yesterday was worth my Dynamus subscription alone. Hey, cool to hear that, Sean. Glad you got a lot out of it. Uh, yeah, our our workshop yesterday was awesome. So, um, Raasmus, he's on my team. I work with him every week. He did basically the entire uh, workshop yesterday. So it's kind of like a guest workshop but like with someone on my team and yeah he crushed it. So basically he gave a master class on creating PRDS and how to go through that planning process with AI coding assistants. He comes from a product management background so man he knows way more about that stuff than me. Like that was a gold mine. I learned a ton from that. So yeah. All right. Um is the spicy mango available at Costco? Um let's see. Fruit Riot. Uh Costco. I don't know, Eric. Um Oh, yeah. It looks like it is [laughter] cool. Yeah, I don't personally shop at Costco, but yeah, I mean, it's I've been there. It's a great store. But uh yeah, anyway, let's uh let's get into the building here. So, let me jump back over to my second brain here. I think I'm going to actually open up a brand new conversation with Claude and I'm going to like I promised paste in this initial build prompt and then I will explain things as it's running just so that we can be efficient here. So there we go. All right. So the first thing is I'm telling it to use the archon skill. So, this is a skill that ships with the archon repository that uh teaches it everything it needs to know about how to run and build archon workflows. And so, that plus uh my instructions here is pretty much everything it needs. And I am being pretty specific here. So, I'm not going to like break down every single line in our prompt, but I'll just describe at a high level what I'm building. And like generally when you're building archon workflows, you don't have to have this like really elaborate. So this is the start of the prompt here. So our goal is to build a generic mix provider workflow that ships a full stack web app by routing each phase to the model that earns its tokens as in like this is the one that we figured out is best for the individual task. All right. So we have the arguments that come into the workflow. So when we invoke this workflow, we're either going to give it a spec of the full application we want to build or just some kind of like free form description, right? In the case of our testing, we're going to be passing in this spec that I already explained for the benchmarking dashboard. So like this is sort of our um PRD plus spec. I mean, it's kind of like a bigger document that I would normally send in, but I just want to like kind of push this workflow to its limits. But it's a pretty big document outlining everything like here's what we're building and here's how we're going to go about building it as well. So it it's yeah like PRD plus spec like everything together in one document here. And then as far as the eight nodes go what I have outlined in the prompt right here is literally everything that I explained at the start of the stream with our diagram right here. So I'm just telling it like I want to have an exploration step at first just using sonnet for the model then going into planning with opus just pretty much verbatim here right explore sonnet survey the repo and spec out the artifact um or sorry load in the spec and then the artifact that it outputs is a context.md. So every single one of these nodes in the workflow is going to communicate to the next node through a markdown document, right? Like this this is the output artifact of every single step. So exploration outputs the context.md and then the plan step is going to read the context. So this kind of sets the stage so it understands the repository and the task and then we can create the more specific plan from that. So like exploration to structure plan and then we output that plan and then we send that into Gemini first to build the UI. So I'm specifically telling it uh when we create the skill for this node or like the prompt for this node, it's going to write out three sections. It's going to write out the UI scope. This is what is sent into Gemini through PI. It's going to have the integration scope. And so that is again for opus like if we're building connections to an API or we're using clerk for authentication for example like I'm doing in this demo right like the goal here is like I just want to show something more than just like a super simple landing page like let's actually build in some kind of integration and have opus handle that because once the UI looks good going back to opus is I think going to lead to better results for sure and then we also have a deployment at the end as well. So each one of these sections is going to be processed by the subsequent nodes. So like build UI reads section A. Um section B is read by the integration and then we have uh sonnet do some validation just make sure that like our tests are passing and everything like that. And uh opus will fix anything that comes up and then we'll deploy at the end. And so like I am going to be adding in the clerk CLI skill specifically because I need that for my example. But aside from like the actual skills that I'm injecting into the node, everything is very general here. Like it doesn't matter what kind of integration you need, like if you're trying to integrate with some other um off service or some other API, like you can do that with this workflow. I'm making it really general. It's just if you need an MCP server or a skill for the integration, not only would that have to be something you mention in your spec, like I talk about the clerk integration and the spec here, but it also is just something where you'd have to like, you know, update the workflow. Like if you guys want to use my workflow that I have that I'm building live right now, you can definitely do that. You can use it to build any front-end application that has any kind of integration. All right, cool. So, uh then I I kind of explained the handoff convention here like every single node has to output a markdown file to the artifact directory. This is a special primitive for archon workflows. Every time we have an archon workflow run, it's going to basically have this working directory that's resolved to the path like this variable resolves to that path. So every single node in archon workflow uh in the archon workflow like points to the same workspace where we can uh pass context between nodes. All right. Um and then yeah just some other like gotchas that I documented for my dry run. Um, again, this kind of thing you just build up with Claude into conversation if you don't have a really detailed spec at first. So, anyway, um, let's see here. So, I specifically asked it to ask me some questions. Just showing a bit of a demo of that. Um, let's see. I'm going to So, it says, uh, how should I scope the tools on the nodes? Um, for the sake of demo here, I'm just going to allow all tools on all nodes. Uh the build UI PI node has real environment caveats you listed. How much how much should I capture as YAML comments? Let's see. Um I don't really know what it means here. I don't think I have to document things in the archon workflow. So I think we're good with that. All right, cool. So I'll answer his questions. I think it'll get right into the actual building now. And I'll pop back over the chat while it does that. All right. [clears throat] Uh, what I usually do is I have Claude do the mockup design on Figma Desktop. Okay, nice. Yeah, there is a Figma MCP uh that controls the Figma desktop so you can visualize the UI before having it work on it in code. Yeah, that's smart. I like that. Um, I mean like maybe maybe Claude plus Figma is going to be better than Gemini 3.5 Flash, but also like what if you used Gemini 3.5 Flash with Figma? Then it might be even better because it's just in general I feel like it's more creative of a large language model. I I think that's like the simplest way to put it is Gemini 3.5 Flash just feels the most creative. Yeah. Um, do you install the MCP in the coding agent like cloud code or in archon? So when you're building an archon workflow, you do actually add the MCP server into the archon workflow like for any node where you want to use the MCP, you attach it. So you kind of like build the configuration into Archon. And if you're curious how to actually do that, then you can just ask your coding agent. If the coding agent loads the archon skill that comes with the repository, it'll know exactly how to uh you know incorporate MCP servers or skills or anything into archon. All right, let's see. Cool. So, it answered or it built the workflow here. It Why don't It doesn't look like it built it from scratch, though. Okay. Well, I guess yeah, so it is it is kind of like so did a backup of the dry run and then the live file is free to rebuild. Okay. So, I guess it's just kind of doing a massive overhaul here, which maybe I kind of want to ask it to just like totally build it from scratch. Um, build Well, I guess Oh, I guess it did. Build from scratch against your prompt only. Validates clean. Okay. Wow. So, that that built pretty fast. Uh so we have eight nodes linear explore plan build UI integrate validate fix validation deploy finalize nice um let's see as our default yep uh looks good artifact handoff every downstream node opens with a step zero read the upstream artifact directory markdown first looks good okay and then we do the deploy with clerk so uh throughout the integration as well So, uh, yeah. Okay. So, we're going to set up our authentication after the UI is built. Yeah, this looks good. Okay. I'm I'm kind of glad this this built really fast because I do want to get right into actually running the workflow. Um, so let's take a look at that here. What's the path that it wrote out to? I want to open this up so we can take a look at it really quick before we run it. Uh, Archon build front end mix. Okay, good. So, let's open that up. um archon workflows um build front end mixed. Okay, let's take a look really fast here. So it's it's pretty long because all the prompts are inline. Uh so we have our exploration. You're gathering context for a brand new full stack web app. The dispatch method message is either a path to a spec or a free form feature description. So like this is the argument that we pass in. So if it's a spec file, read it and summarize. Otherwise, if it's free form description, summarize what's being asked for instead. So we're writing a one-page context dump concrete file path versions, what's present, what's missing. Really just setting the stage for our build. And that goes into now opus for our planning here. So you're planning a full stack web app. Before anything else, use the read tool to open context.md, right? Like this is the file that the prior node outputs. That's how we're communicating between our nodes here. So in in Archon, there are two ways to communicate between nodes. Either you have them write out to a file like this or in subsequent nodes, you can just reference the ID of the node.output. So like explore.output is going to be whatever it output in the last turn from the large language model in that node. But just to be more comprehensive, I want it I just want to like output to markdown documents because it also like gives us a record of context as the workflow runs. So we can even like view these files as the workflow is running or after the fact. So this is where we're going to produce the plan with three sections. UI scope for Gemini and then integration scope and deployment plan. And then here's where we uh use PI with Gemini 3.5 flash through open router. You're building the UI um only the UI surface. And here's the important part. O API calls third party SDKs. It belongs to later nodes. This is just the interface. So we read the plan and we're telling it that the work you're supposed to work only from everything under section A UI scope. So, we're telling it here to ignore section B and C. And honestly, if I wanted to make this workflow even tighter, what I could do is I could have a deterministic step in between these two nodes that would like automatically split out the plan into a section A, B, and C and then just like load only that as context into this second node. So, like here, I don't think it's that much of a problem. I I'm sure it'll be fine when I'm asking it uh to by itself like take that part out of the plan, but like if your plans got really long, it would actually be quite beneficial to never even have it load the other sections, right? So this is just more for simplicity sake. Like let me just shove everything in one file, tell every node to read the entire file but only like care about one section. This is going to work here, but definitely some ways I can improve this workflow. So it's just kind of a balance of like simplicity and speed for the stream versus like how comprehensive is it actually going to be. And then for the integration, this is where we are using Opus again with the clerk CLI skill. Um so yeah, like I specify here, the clerk CLI skill is loaded as a capability. Use it only if section B picked clerk for authentication. So this is a general workflow. Like if you don't have any kind of authentication at all, like it's just going to ignore that part. But I'm just building it, you know, a little bit of fit for purpose for today. All right. So then we output the integration summary. We do our validation blah blah fix validation. Don't need to go through all this right now. Then deployment as well. Um, so we'll use the clerk CLI again for that. I mean, we're using it through this entire thing here, which is pretty cool. like we're going to have it's kind of crazy how difficult it used to be to create authentication in an application and now like with coding agents and clerk like shipping this skill using their CLI you set up everything so easily. Um it's it's very cool. So like we'll we'll see that all in action as well. So let me jump back over uh to Claude here. Scroll down and uh let me copy the path to the spec here. So I'll just copy this path from Obsidian. I'll paste this in and I'll say all right this is the spec for the application I want to build. I want you to now dispatch the Archon workflow that we just built with this path as the argument so that uh we're going through the full front-end build with this spec for the Archon benchmarking dashboard with clerk authentication. Uh let's set it all up. All right, cool. So, there we go. We had the the rainbow color for workflow again. Hope it doesn't actually build I hope it doesn't actually use the sub agents for this. I I really don't want I just want it to kick off the archon workflow and that's it. And then the other thing is it would be nice to actually have the UI up and running. So let me do that here so that we can monitor this. So I'm going to say um yeah load the archon skill and then uh I want you to start the front end and back end of archon for me so we can have it handle this for us just so that I can have the web UI up to so we can watch this workflow as it's running. So, I kicked it off right here. And uh now we just have to wait. But I want to be able to see like where where we're at in the process. Okay. Uh one important call before I launch the build UI node runs PI and open router which needs the patched archon source. Compiled Archon binary still ships. Um okay, sure. That's fine. Yeah. So that it's I have the wrong version of the binary, but Claude caught that and fixed it, which is uh pretty nice. So all right, great. And then after the archon workflow kicks off and we have an artifact directory, I want you to share the artifact directory with me so I can watch the markdown documents come in as the nodes complete. All right, cool. So that way we can kind of watch the context get formed as it's going through the full thing because it'd be cool to like kind of see the website form like what's the UI look like before authentication? What does it look like once we actually have clerk set up? Uh, okay. Okay. So, work tree created. Looks like it's working. Okay. Find artifact. All right. Nice. Yeah, we're looking good. All right. Did the stream go down? I hope not. I I think I think we're still here. This guy again dropping the big boy stuff. [laughter] Mixing providers is the big boy stuff. Yeah. Well, happy to be doing it. Uh, all right. Oh, a lot of you guys have actually lost the stream. That's a bummer. Can you guys confirm? Is the stream working right now? Because I see a lot of you are still here. Maybe there was just a blip for a little bit. Apologize, guys. I don't know. I don't know what uh would have caused that. busy preparing the next banger. Beautiful frontends part two. I am planning on making a YouTube video based around this stuff after the stream for sure. Okay, there was a blip. Oh, that's a bummer. Um, okay. It's working now. Okay. Uh, it's working, but it went down for a while. Shoot. Sorry. I have no idea why that would happen. That's a bummer. I did see like the viewership drop for a bit and I was kind of confused as to why, but yeah. Shoot. Okay. Well, I think we're back now. I I really don't know. Like there was no indication on my end. Um so yeah, I apologize guys if uh there's anything that you missed there. Like if I because I might have been in the middle of explaining part of the prompt or something. So hope there's nothing critical. I honestly I'll have to go through the a recording after and see like if the YouTube live like the recording posted after actually has all the content or if like the blip affected that too. I'd be pretty interested. All right. Yeah. Down. Okay. Down for a couple of minutes. Okay. At least it's only a couple minutes. All right. So, good. Pay attention viewers because this practice in front end can teach you a lot about how to work in your agentic setup. Well, I appreciate it. I'm glad this this workflow is looking good to you. Yeah. I mean, the the most important thing here is the context between the different nodes. And while we're waiting for this to run, one important thing that I want to explain here is like not only is it powerful to have a lot of different steps so that we can use different models, but it's also so that we just get better results overall because large language models, they get overwhelmed just like people do. If you try to have them do too much at once, they're going to fall flat on their face. And so you always get the best results no matter how good the model is if you make it very focused on a specific task. Like your job here is just exploration or just planning or just building the UI. So even if I was using Opus for everything, having this workflow is still better than taking this entire thing and trying to jam it into a single prompt or a single skill. It's not going to get the best results. Now, obviously for a more simple front end uh with simpler integrations, Opus probably could handle the entire thing in a single prompt, but I'm interested in building workflows that can really scale the complexity of what we're actually building. Like that's one of the reasons why I want to use Clerk for example, so that we have like not that like Clerk is fundamentally difficult to set up or anything, but just like having like a real integration for like full authentication built in just proves that like we're making more than a dinky landing page. And in fact, this application is actually pretty complicated overall. This entire benchmarking dashboard, I I don't think it's going to be able to oneshot it. But it's going to get a lot closer with this workflow than if I just like threw the entire spec that I showed you earlier into a single cloud code session. That's not going to end well. All right, cool. So, let let's see where we're at now. Back over. Okay. Uh here's our artifact directory. Very cool. Uh I'm going to open this up in a new VS Code instance here so we can look. So looks like the first node completed. [snorts] So we have our context.md. It is an empty scaffold. There is nothing here. We're starting completely from scratch. And so really like in the case here, the exploration isn't really finding much. It's more just kind of specking some things out initially. So we're going to create a next.js 15 app. Uh we're going to be using clerk for our authentication. And so loading the clerk skill, it's going to know how to like initialize things and and get it configured into our front end. And uh what else do we have here? Our archon integration. Honestly, it doesn't really need this information. So maybe not the most ideal context, but that's fine. uh because whatever it passes into the next node then then it'll start doing the planning for the actual application. So I guess in this case it's acceptable right like this is just our highle exploration of everything that we got going on here and so the workflow is running now um on the second step. So let me jump back over make sure our UI is good. Yep. Okay. So uh let's open that up here. Let's take a look at the workflow. So, I'll just do localhost 5173 for Archon. All right. So, here we are. So, if we click into this, we finished our first node [snorts] for our exploration, and now we're just in the middle of planning here. Cool. All right. Um, oh, and then one one kind of fun thing while we're waiting for this, there's a little bit of a Easter egg in the the clerk CLI. So, um, clerk, I thought I thought they had a d-help. So, there there's, uh, yeah, here it is. So, a [snorts] lot of these things like the initialization and the linking and the deployment and everything like we're going to have our Claude code workflow or our archon workflow handle all of this with Claude. So, we're not really going to run any of these oursel, but they they added in an an Easter egg to to Clerk. Um, I think it's clerk bird. Yeah. Look at this. We can play Flappy Bird in our terminal here with the clerk CLI. It's the most random Easter egg, but it's kind of fun. So, um, yeah, I'm I'm going to play a little bit Flappy Bird while we wait for the Archon workflow. I planned this, by the way. Like, I I knew they had this Easter egg and I wanted to do this, and this is the perfect time because we're waiting for the Archon workflow to finish or at least get to the next stage so we can take a look more at what it's planning. But yeah, so if you guys want to install the Clerk CLI, I got a link to that in the description and you can you can play Flappy Bird, which is pretty cool cuz it you can't play the official Flappy Bird anymore, right? They took it off the app store like, you know, a million years ago. It probably been like a decade, honestly. I was never good. Like, I'm probably going to suck right now. But also, I wouldn't want this to go for a million years anyway. It's a little laggy, but overall, it's pretty good. Oh, yeah. Okay. Anyway, there there's there's a fun little Easter egg since we're using Clerk for our workflow here. Okay. Hope the live stream's not dropping again. Um, doesn't look like it is. I'm I'm just kind of like going to be anal about that now and watch it constantly. [laughter] All right. Um, is Figma necessary now that we have clawed design? That's a good question. So I personally haven't done a lot with Figma. Claw design is really good though. So I think it depends like for people that already have Figma integrated into their their workflow and like they have a process for using it. It's probably going to be better. Like I don't think claw design is like inherently superior. So you'd want to switch over. But like if if you are more interested like getting into design with coding agents for the first time, claw design probably would be easier to start using. All right. Uh, miss the days where you could actually learn from YouTube just by listening to a creator share their knowledge. These days, it feels like it's just about reading AI generated text. Yeah, I I understand, man. Uh, I I hope that for this live stream, like, yeah, I'm reading some AI generated text because I'm going through building a workflow with Claude, but I'm still trying to stay really like meta, like helping you understand like here's how I'm working with Claude. like I'm reading I'm reading off what it's saying to give context of like the larger explanation that I'm giving. It's not like the entire lesson in this stream is just reading off the text, right? Like I'm explaining the the whole like diagram that uh or yeah like the whole workflow in the diagram that I'm building out here and like this isn't just like AI made all this up, right? Like this is my workflow that I designed. I obviously worked with Claude to help me design the diagram and ideulate, but like this is my idea that like now I'm sharing with you like everything that goes into this workflow and exactly how we're handing off context and stuff and the specific model that I'm choosing for each step like this is all me like entire like Claude did not suggest any of this. Now, I can understand how in doing a lot of prep before the stream, it might kind of feel like I'm just like presenting a bunch of stuff that Claude threw at me, but like trust me, I put hours and hours into preparing a lot of this. So, I hope that makes sense. But I definitely understand where you're coming from. But yeah, whenever I do a live stream or a YouTube video, it is always like there is some LLM generated stuff like you know like it helps me plan like it helps me take my ideas and put it into words and diagrams but it is still like my genuine ideas but like I'm also with you like when I first started my YouTube channel I did not use AI coding assistants at all like every single line of code was written by myself Okay. Um, let me sorry. Let me pull up something really quick. I'm going to pull up my YouTube channel fast so I can show you some of the videos that I used to create way back in the day. We're going to go uh through a trip down memory lane just really fast. I mean, we're waiting for the workflow to run right now. Um, so like we got some time. So, let's let's go down memory lane for a second here. I'm going to go all the way to let's do like we're going to go to July 2024. Let's go to July 2024. The thumbnails get progressively worse as as we go down here. Um oh boy. [laughter] All right. Um so the very first video that I created, uh wow, this actually has 16,000 views. Holy cow. So, when I was like first getting into YouTube big time, like obviously I have some like kind of other stuff before, but I wasn't really like trying for much then um was my AI agent master class series that I did. And uh wow, this there are some silly videos here. But anyway, like in this video, like let me let me open up one of these really quick. I'm I'm just teaching like here's how to build AI agents. And at the time, I was using Langchain and like I I wrote every line of code here completely by hand. Um, and like yeah, this is a pretty basic agent, but like then I get more into like langraph and more complicated agents with pinantic AI later on and and lang like I'm I'm writing everything by hand. I didn't I didn't start using AI coding assistants until December of 2024. So there was like a good like four or five months of YouTube content where everything you saw was completely written by hand. But the the problem is it's just not realistic to do that anymore. Like the thing I want to teach is leveraging AI coding assistance. So, I better be using it in everything that I'm preparing for you guys and in the actual demonstrations. But I am still always applying my teaching on top of it. Like it's not just me regurgitating LLM generated stuff. So, I mean maybe a bit definitely longer answer than you were expecting, but I thought I thought it's just worth explaining that like and also kind of cool to talk about the the journey of like it is crazy to think that back then like in this video um back when uh I can't uh go full screen so you can't see my face, but I used to have like way too much space um at the top of my camera there. Oh, good times. Good times. But anyway, uh, all right, let's see. If that's criticism, then it makes no sense. Like saying someone explaining how to use an app is someone reading the app. Um, yeah, I mean, I agree. Yeah, but like I mean I I I don't Yeah, like I I agree it's not the best criticism, but like like I said, like I understand to an extent, so I'm trying to like Yeah, kind of bridge the gap there. Anyway, Sean said, "What what where are your Minecraft videos? Thought you would repost my [laughter] Damn it, Sean." No, I'm kidding. Uh, yeah. So, that's something I told the Dynamis community is uh I this YouTube channel that I post on like now I did actually use as a kid. I just unlisted all those videos, but I didn't actually create a new channel when I started my AI content. So, this is like my original YouTube channel that I created when I was like 12 years old or like 11 years old. Yeah, I l I created my YouTube channel when I was 11 years old. Um, so yeah, I had some Minecraft videos. I like every other kid on the entire planet, I had some Minecraft videos on YouTube. They they are just unlisted now, but theoretically, I could play one on the live stream here. I'm just I'm not going to do that right now. That [laughter] would Oh, I would die of embarrassment. That would not be good. Oh man. All right. Uh let's see where we're at now. So I can say give me a status update and uh see if we progressed and then I can check the logs as well. All right. Let's see where we're at. All right. Oops. V logs. Okay. So we're building the UI now. Okay. Cool. So we did all of our planning. We're building the UI. Looks like we have some tool failures though. So there might be something going on there. Unfortunately, I might be hitting a rate limit with open router actually. Uh that I need to check that quick. I might have to add credits actually. Um let me check. Okay, I I can just show this because I'm not like going to show an API key or something. If I look at my activity, um, look at my credits. No, I have credits. I'm good. But it might be a limit on the API key itself. Uh, no, I'm good. I'm good. Okay. It might just be random failures that cuz Gemini 3.5 Flash I don't think is the most reliable in general. I mean, this kind of goes with what I was saying, like besides how creative it is, it's not the best at reasoning. So, it might uh Yeah. So, there's eight bash fails on this uh on this node. They're not crashes. The pies harness flags certain shell commands is danger. Oh, okay. Okay. So, this is just something with pi. So, I think we're actually okay. So, yep, the build UI is running. Okay. So, we're actually good. So I maybe it's not actually a Gemini problem. I think it might just be a pie warning that's just showing up kind of weird in Archon. There's probably something there that we could make better in Archon. So that that is like presented better. It doesn't cuz it kind of looked like the model was just screwing up a bunch, but it wasn't actually the case. Anyway, um hey, don't just stop in the middle of your origin story. What else do you guys want to know? I felt like I kind of gave the story it. Seriously, if you guys are curious about like anything else, like how I got started, like let me know. I just I just don't really know like what I would continue there. Just talking about my YouTube. Um, [laughter] so yeah, my YouTube channel, aside from Minecraft, uh, as a kid, I also posted Overwatch videos from when I was in high school. So, uh, in I don't know how many of you guys play video games or or have like played Overwatch, but I got way way into Overwatch in high school. Um, I got to I actually got to number 218 in North America. So, I had like, you know, the the big top 500 status that's kind of in like a lot of different ranked video games. Uh, yeah, that was that was a lot. And then I I went to college and um I I was still playing Overwatch at the very beginning of college. But then the problem is um I mean you guys none of you probably even care, but the the like best characters in Overwatch like the meta as it's called became this like really really just the game the game fell apart in um 2019. So I I stopped playing. Uh yeah, but but uh I mean actually more recently the game is kind of fun again doing playing it like a little bit. [laughter] Let's see. So uh all right, what are the pros and again let me know if there's anything more you guys want me to like talk about. I'm I'm I'm open book. Uh what are the pros and cons of Archon over Claude Code's dynamic workflow? So that's a good question. Um, Claude Code's dynamic workflow though is is not really the same thing as Archon at all because it's more about just orchestrating sub agents. It's not really defining full workflows where you're mixing providers and you're passing context between them. There's not much that's like really sequential about it. I think it's more just like orchestrating in parallel. I'd have to dig into it more. But like in the end, if Claude Code ever builds something that actually is like Archon, which they did have kind of like a redacted feature, they they their change log published something that did look more like Archon and then they took it away. It's probably coming at some point. But the thing is, no matter what claude code ends up building, the moat that we have with Archon is that it's open- source and that you can mix providers. Like exactly what I'm doing in this stream here, mixing Gemini and Claude. Like if you're if you're working on a harness in cloud code, you're forced to use cloud code. And that's something that a lot of people are not okay with right now because of how expensive cloud code is, how much the subscription is getting worse, and the API costs are ridiculous. like I'm I'm specifically interested in using archon so that I can you know use pi for Kimmy or Gemini in certain steps and then still rely on cloud code because I think opus is the best but if I want to test with GBT 5.5 then I can do that as well I can use codeex and archon too and um yeah there there's just there's so much work that has gone into building archon where like claude they got to make it good to make it as good as archon there's so much that we handle with work trees and stability with that and like the artifact directory and how you can inject skills and the different parameters that you can tune in archon and you the fact that it is mix provider like there's so many things that make it so that like I I don't I genuinely don't know of another tool that that is um close to archon being able to build any kind of harness that you want and build it like not not vendor locked right so yeah I think um let me actually cloud code dynamic workflow. Maybe we can take a look at this really quick like what it actually is. But yeah, really it's just orchestrating sub aents, right? Like that that is I mean sub agent orchestration is a harness but this isn't like a harness for anything like what archon can be like archon allows you to build any kind of harness. So a dynamic workflow is a JavaScript script that orchestrates sub aents at scale. Claude writes the script for the task you describe and a runtime executes it in the background while your session stays responsive which honestly doesn't feel that different from what we could do with sub aent orchestration already. I think it's more the just the fact that like it can run it in the background honestly like it it's really not like archon at all. I mean archon yes is also workflows that you run in the background but that's for like an entire AI coding process like exploration and plan then implement then validate and another thing that this doesn't have is any kind of determinism right like one thing you want to do in archon sometimes is you want to run a command like after the implementation you run the tests or before you do your planning you load in some context deterministically like that there's none of that in here as well so you're you're really still just like shoving your entire system in the agents instead of building the agents into the system like applying the the reasoning from the large language model where you actually need it. I still think this is a good feature. Like it's really great. I've used dynamic workflows a couple of times now, mostly just for like research with sub agents. Um but it is it is really nice just being able to dispatch it and then continue a conversation and then once the workflow finishes then you can kind of like go back to that in the conversation. It's nice for sure. Yeah. All right. Cool. Let's uh go back to Archon and see where we're at now. Dashboard, view logs. Um Oh, we've actually built the UI already. Okay. Uh cool. So, we could actually check it out now. I'll wait a little bit. I I'll go back to the chat um for some time here, but we could start the UI now. I just don't know if it'd work yet because we haven't actually gotten to the validate step. But yeah, I can say give another status update here. Just see if there's any other issues that have come up. But I think we're looking good, excuse me. Um, am I using clerk for off now for everything instead of superbase? So they serve different purposes really. Excuse me. You It's not like you have to compare them directly. Superbase is simpler authentication, but it's not as robust. Clerk is like the enterprise scale or at least the the authentication that can have enterprise scale if you want to reach that. There's a lot more features around user management. Honestly, it's easier. It's even easier than Superbase to build in now because of their CLI and skill and just how much you can like have your coding agent build and like even deploy and set up the entire environment for you like we're literally doing here. And so that that's why I'm using it right now is just because of how integrated it is with coding agents with the CLI skill. It makes for a much better demonstration in the workflow here. So I actually reached out to them to to work on on this. So I'm like, "Hey guys, I want to I want to incorporate clerk into this here. Let's get some authentication in the application." Yeah. But yeah, Superbase is like um so okay, let me let me clarify. Superbase is simpler like in order for you to like manage it in general but like and it's like just more lightweight I guess but yeah clerk is really comprehensive like that it is like the enterprise level solution for authentication okay [clears throat] anyway so uh 3.5 out of eight nodes all right integrate is running uh Gemini built a real Nex.js app. Uh, the build UI finished with a closing message, but it never wrote the UI summary into the artifact directory. Oh, interesting. Okay. Um, integrate is opus and the real signal stub files is sitting right there. Ah, see this is exactly why I don't want to rely on Gemini for the entire thing. It literally missed an a huge step of the prompt that we gave it. Like if we if we go look at the workflow right now and we look at after plan we have the build UI. We are specifically telling it to write the UI summary listing every file created and integration stubs in the codebase for where we need opus to integrate the authentication or APIs or anything. And it skipped this it there there isn't actually we could actually verify. Yeah, take a look at this. the plan or the build UI finished and it didn't write out the markdown document. So, Gemini 3.5 Flash completely missed a part of the instructions here, which is a big bummer. But also, it did actually build the the UI. So, like we have the UI, the workflow is not screwed, but uh yeah, this just goes to show that like it's not the best model overall. It's a creative model, but it's really not the best. Um, so yeah, I'll have it alert me. Uh, keep watching the integrate. Ping me the moment it lands or trips on the missing file. It shouldn't trip. Hopefully, it says that we're probably good, but that is good to keep in mind. That actually is quite the bummer. Um, yeah, like definitely I don't think Gemini 3.5 Flash is even as good as Sonnet as far as reasoning goes. Like I don't even think Sonnet would have missed that obvious instruction there. All right, let's see. Keep up with all your excellent teaching skills and material. I'll watch offline. Sounds good. I appreciate it. Haters going to hate. Um I appreciate it, Sean. Good to see you, by the way, Sean. Um you know, the other comment from earlier. I wouldn't even call it a hater. Like I think that they have genuine um concern. I don't even think it's necessarily hate. I I guess I don't know the intention for sure, but still. Anyway, I I appreciate it, Sean. Of all the creators out there, you do break things down into lessons that are easy to understand. The process process matters. Yeah, exactly. I I just want to teach the process. And sometimes teaching the process does include reading off what the LLM spit back at me, but that's just how it has to be. Uh, what made me quit my full-time job to do Dynamus? So, I actually quit my job before Dynamus. I'll give you guys a little bit of a timeline here as we're we're waiting for the workflow to finish. So, um I I worked for uh a wealth advisory company, financial. I mean, it's public information. You can go look at my LinkedIn. So, I don't I don't need to like be secret about the company or anything. I worked for Marrize as a software engineer. So I was a web developer and cloud engineer and um also did some like AI proof of concepts for the company for like my last year of working there. And um I started my YouTube channel while I was still working there. So to get specific with you, I started my channel like super end of June or start of July 2024 while I was still working on air prize. And I told myself, I'm just going to kind of do it on the side because I just wanted to do it for fun initially cuz I was like helping my friends build agents for their startups. Like I said, I was doing some stuff for prize. So I I was I like had a lot of that um expertise size building up really quickly. Not that I had a ton of expertise, but it was still like more than everyone else because it was just so new for everyone. And so like I like, all right, I'm going to start a YouTube channel. Just start teaching some stuff cuz I I actually really enjoy making videos and teaching. And like I was a TA in college. I was a chess coach in high school. Like I I like teaching. And like as an individual contributor at a company, I didn't get to do enough of that. So I'm like that's why I want to do a YouTube channel. I didn't think I would really get much of an income from it. Um because at the time I didn't really like know how sponsorships worked and and um I didn't really like think that I'd have a ton of people reach out for consulting and stuff. But like all that started happening and so it got to the point in in September, so like literally like within 3 months of starting the channel, July, uh August, September, that uh it replaced my income or like it equaled my income at at Marri. So I'm like, you know what, this is what I really want to do. like the the job was great, but it wasn't like my true passion, right? Sort of like a like this is good but not amazing kind of thing. Um, so I I left and I started doing YouTube full-time. And then uh Dynamus actually uh really like the idea was birthed like January 2025 and then I worked my butt off with the initial modules for the first course. Like oh my gosh, I I grinded that out and then released the community in April. And so yeah, the Dynamus has been around for about 13 months now and it is like the best thing ever. So yeah, that that's kind of like the the timeline there. Yeah. And like as far as like why I quit, I guess I kind of already said that, but it really is just like teaching and generative AI are like my two biggest passions. And okay, honestly, I will say that uh education, I'm probably more passionate about education in general than I am like actual AI. Like of course I'm like insanely passionate about AI as well, but if there's like one skill in the entire world that I could be the best at the best at, it would just be teaching. Like even more than like building or using AI like it teaching is like the biggest thing for me. All right. Uh cool. So anyway, I'll keep going through what you guys got in the chat, but I just want to check really quick on the status here. Um okay, still integrating. Yeah. So, we're we're getting through it. All right. I make you feel old. Well, you're not old. We're all young, right? [laughter] I got in on it very I got in on generative AI very early. I I would say that a lot of like younger guys like myself and and and like Nate Herk, for example, like we got in on it really early because we're just starting to like, you know, figure out our career. like what is the thing that we're going to dive all in on? Um, so like kind of before generative AI, I was sort of the jack of all trades, right? I did a lot of like cloud engineering, a lot of web development, systems building, like and so I always knew though like that there's one thing that I want to focus on at some point. Like I knew that like you know niching down as they call it is pretty important. and uh didn't really know what that would be until Generative AI became a thing, you know, with like chat GPT uh 3 point or GPT 3.5 Turbo at the end of 2022. That was like when I knew like, okay, November or December of 2022, like this is going to be a big deal. And then especially like come 2023, that's when I was like building agents and everything all the time. uh especially towards the end of the year and then getting into the start of 2024 is when I really started like building those skills that led to me creating my YouTube channel. Um you loved Overwatch 2. What about Marvel Rivals? I've never played Marvel Rivals before, but yeah, Overwatch is great. I I do actually I have a friend who is a professional Marvel's Rival player though. my um old uh not old friend, he he's a current friend, but he doesn't live in in Minnesota anymore, but uh if anyone is big into Marvel Rivals, you probably know Kova, but uh that that dude was like he was over at my house like every weekend in high school. We were playing Overwatch so much. So yeah, good friend. And uh yeah, so I don't think he he'll like ever tune into a live stream, but if you're here, Mark, cool to see you, man. All right. Um, and then how did you join the dark side? There must be something traumatic to stream Python. You got to elaborate. I don't know what you mean. [laughter] Uh, he said you're an open book. Please, can I have your open router API keys? Um, sorry. I'm an open book for my timeline, but not my open router API keys. Sorry. Sorry, man. My second brain won't let me give that to you. [laughter] All right. Um, okay. Just to clarify, my criticism wasn't aimed at Cole, but the current state of YouTube in general. At least Cole actually shows the workflow. Okay, I appreciate it. See, I knew that like you probably weren't just being a hater. Um, and like and again I I see where you're coming from and I think like I try my hardest to not just regurgitate LLM stuff and like I say I try my hardest because it is kind of tough like sometimes it is tempting to be like you know okay uh second brain here's this article from Opus 4.8 go ahead and create a YouTube script and then you just like present that and you're not really even applying your own original thought at all. Like I I know that like a lot of people do that because it's easy. Not saying everyone who covered Opus 4.8 just regurgitated stuff. Um a lot of a lot of creators I I think do a really good job at like really like finding some interesting angle that they're presenting their own thoughts on, but still it is tempting overall. So like yes, there is a lot of that on YouTube. I totally agree. Uh, my recommended feed is just full of videos that just tell you which repo to clone or download skills from. Honestly, feels like a majority of them don't even know what they're doing. Yeah, there there are a lot of channels like that where like their entire model for the channel is like just share some kind of AI coding thing that has viral potential, right? Like with superpowers or GSD or BMAD. It's like whatever that new framework is. It's like, "All right, make a video. Make a video. Make a video." And and yeah, I I would agree that like that's not something where they're really applying too much original thought. Like those channels are still interesting to follow because it is cool to uh have those things surfaced where it's like, okay, may maybe I do actually want to take a look at this new uh AI coding framework and see if there are some good ideas for me to build into my own. My general recommendation is usually not to run any of these uh or like use any of these frameworks yourself because they're very opinionated and generally bloated, but there's still a lot of good ideas that you can identify from them or like maybe there is like an individual skill that you'd want to incorporate. So like there's value there if and like especially if you don't have your own kind of like second brain system that like finds those things for you as they're trending on GitHub or whatever, those videos can be like the way for you to find out about it. Um, but yeah, it's not like they're really teaching anything. I agree with that. All right. Uh, Ivan asked, "How will Archon work once Anthropic changes the agent SDK costs? Also, the second brain. Am I reverting from the agent SDK in favor of something else?" So yeah, that is an unfortunate thing where starting June 15th, what's going to happen with Claude is there is going to be a separate $200 a month credit that you get for using Claude programmatically through the agent SDK or through, you know, like Claude-P where you can invoke it in headless mode. You're not going to be able to use your Enthropic subscription anymore. So, it's still nice that you get like $200 a month that you get for the API credits when you're using Claude with the agent SDK. U but that $200 is going to go very very fast. And so, as far as Archon workflows go, um this is a great question because this is one of the reasons I'm so interested in mixing providers right now because I still might want to use Opus and that $200 credit for some of like the planning that I'm doing in my workflows, but then for a lot of other things, just leaning more on cheaper models. So, that's one thing I'm exploring um because it it's definitely like this change is going to affect Archon quite a bit. Um same thing with the second brain. The other thing that uh I'm going to probably be leaning on is the codeex SDK because um OpenAI is much more lenient for how you can use a subscription. Way less uh ways to violate their terms of service. You can use it programmatically still. And so I am actually considering migrating my second brain to use the codeex SDK potentially. The second brain isn't nearly as expensive as any kind of like actual AI coding with Archon workflows or whatever. So the $200 a month credit is actually probably enough for my second brain. So I might keep it on the agent SDK. But that is something that I will uh be spending a lot of time on just experimenting with in in June like once we hit this limit here. like I'm I'm basically going to wait until I have to to like test things out. Uh but yeah, codeex is always an option and then just generally mixing providers so you're not relying on uh Claude as much overall and like like I proved in the last couple of live streams, you really don't have to. All right, let's see where we're at now. Uh it's still integrating here. Um okay, hold on. I'm going to actually take a second to read this and see what's happening. It says that the integration is working well. Uh, another status update, please. It's taking a bit longer than I thought it would. Um, I guess like I just want to make sure it's not stuck on some like clerk um command that's like interactive or something. Let's see. Check whether it's actively writing or gone quiet. How long it's been, we'll see here. I'll check the context files as well because one thing that'd be interesting to read is the actual plan here. So here is our UI scope. This is what Gemini 3.5 Flash took. Very cool. Um may honestly it probably kind of described the UI a bit more than it should have because the point is for Opus to decide the information not necessarily like how it takes form. But hopefully Gemini 3.5 Flash is able to run with this and make it look good. And then uh let's see. Let's go down to the integration. Yeah, here we go. [clears throat] Okay, so authentication with the clerk. We're going to have it run the init. Um and it knows that because of the skill. It's going to set up environment variables and then create the route handler for authentication as well. And then it's going to set up the API endpoint. So there's actually a lot that's doing with authentication and the API as well. That that's actually probably why it's taking so long. I for I totally forgot like integration is much more than just clock clerk for the oh my goodness Clark. It's way more than just clerk for the authentication. It also has to set up all the APIs to work with archon because this is a benchmarking tool that we're building here. Uh so yeah, sorry I forgot that this is like a massive application. It actually makes sense that it's taking this long. Okay. And then uh the last one is the deployment section C deployment plan. So we're going to deploy to Verscell and then we're going to run the clerk deploy to promote the clerk instance from dev to prod uh versel serverless functions got host the archon cli prod. Yeah, that's that's kind of a tough I like we're not really going to have like a final production deployment here because I can't deploy Archon, but I more just want to show off like getting the user interface out there and and kind of finalizing the authentication setup with clerk deploy just so you can see like like with the clerk skill, it's able to do everything completely end to end like nothing set up in clerk all the way to uh dev for testing prod flip. It's Yeah, it's insane. Okay. So, yeah, the plan the plan is solid. So, yeah, we'll just have to wait for it to get through the integration here. All right. Um, integrates still running healthy, actively writing. Uh, yeah, it's it's looking good. It's now in the clerk o data layer. Uh, good. Still pending. No middleware yet. Um, no new bash failures. Yeah. So, okay, we're looking really good. Um, let's see. So, there hasn't been a file, right, for a while now. Let's see. So, I'm not going to call it healthy. It's watch it closely moment. Okay. Watch watch it closely moment. The work so far is real and substantial. So, even if integrate timed out, fix validation, a rerun could recover. Okay. I mean, I I think that it's just um thinking deeply right now. So, we're looking good. Good. We're looking good. I'll go back to the chat here. [snorts] All right. Yeah, thank you for clarifying. I agree. There are many creators which literally read the repo. It's so annoying. Yeah, there is a lot of that. The AI unfortunately the content creator AI space is just super saturated right now. Uh because I mean it is a great opportunity. It's like people get away with just reading the repo. like there there is a huge audience that just eats that up even though they are really just kind of you know regurgitating and I'm not going to call out anyone specifically and like I said I think most creators don't do this but there there are still a good chunk where they are just like reading the repo. Yeah. All right. All right. Tell them about Perdue. I don't know what you mean by like Purdue University. I didn't go to Purdue. That's the thing, right? Purdue univer I did not go to Purdue University. Yeah. I'm not sorry. I don't I don't know what you mean. All right. Um, Gemini is incredibly intelligent, but it's like having a schizophrenic. Is that why it didn't write out my markdown document? Uh, I never use Gemini for anything. I mean, mostly same, but like after 3.5 Flash and the UIs it can make, like I'm interested in using it right now. That's why we're testing it here. Out of nowhere, it just does the weirdest things. Anti-gravity is a mess. Yeah, fair enough. Yep. Um, any plans on trying out Grock Build or integrating into Archon? Um, so I mean I'd be interested in trying Grock Build, but I I haven't tried it yet. And as far as integrating it into Archon, do they have an SDK? Because they they would uh need that for it to be possible. Let's see here. Um, [sighs] I don't think they do. Groc build SDK. It's an interesting new tool. Like I definitely want to try it out, but yeah, I don't think they have an SDK. Uh to build applications, developers, you can try. Oh, I guess so. You can use Grock build within the AI SDK. So maybe it's possible. But that's kind of weird though because like I thought Grock build is the harness around Grock. So how is it so that it's a model that you can use in the AISD? That doesn't make sense to me. I would have to look into that more. I've never been a huge fan of Grock in general. So I'm not like super interested incorporating into Archon, but if there ever was a someone made a poll request to add it as a community uh provider, that'd be great. I'd love that. Uh, created a CLI for my Asian box CLI with the Archon workflow. That's cool, Jeffrey. Nice. Yeah, Jeffrey, you're always up to some pretty cool things. I'd love to see that. Nice. Um, Code Rabbit, how do I feel about it? And folder agents. What I mean by this is running the code with no UI. So, uh, Code Rabbit, I think, is pretty great, especially because it's free for open source projects. So, I've used it a lot for like Archon and and Bolt other other projects in the past. Um, I think that like you have to do a lot of customization to really make it work as you want. Otherwise, the reviews are very verbose and kind of hard to understand and they're like a little um silly sometimes with their output and like so yeah. Anyway, you kind of have to customize it, but it's good. Um, I know like a lot of people really like GPL as well. So that's another really solid competitor that some people think is better. Like there's never it's kind of a debate like what tool is best. Also like building your own just like using cloud code in GitHub actions for example. Just having it in your own CI is also really good for code reviews. So you can like have a GitHub action that triggers whenever there's a new PR and it automatically does the review with cloud code. Like you can you can kind of make your own code rabbit really easily honestly. So, like I don't I don't find those tools like insanely valuable, but they are nice and like especially because Code Rabbit's free for open source. It's like I use it. Um, and as far as what you mean by running code with no UI, I mean, isn't that's kind of what we're always doing with like clawed code in a terminal, right? Or do you just mean that like there's not as much visibility? Like I I'm only going to be if I don't have visibility and it's like through a GitHub action or code rabbit or something. If it's code review, I'm good with that as long as it ends with a comment on a pull request. But like generally, I want my coding to be interactive. Like even in my archon workflows, adding in human in the loop. All right, cool. So, let me go back and see where we're at now. Um, oh, the integration is done. Uh, okay. Oh, wow. Nice. All right. We're on our validation now. So, we also have the plan or the integration summary here. Um, okay. So, clerk, we have our dev instance up. This is so cool. Like, we don't even have to set up API keys cuz the clerk CLI with the skill goes through everything with us. This is so cool. Okay, so here are all the files that we edited here. So, all the stubs that the that Gemini 3.5 flash left for implementation or like for the integration, we took care of that. We created all these file. There's a lot of work that was done here. That's actually pretty crazy. Okay. Uh we logged in with my email. That's super cool. So, wow. It it like this this is crazy how much work this kind of thing would have taken before. Like setting up authentication, testing authentication, going through the whole application. Like all of this, we just blitz through it as I'm chatting with you guys in the stream here. Uh this is nice. Yep. And the MPX build was successful as well. But we got like full proof of everything like we ran the clerk doctor to make sure everything is set up right. Um production not so the one warning we got is that production is not configured but that's what we run with clerk deploy in the deployment step of the workflow. So like everything is literally perfect here. Okay. Nice. All right. So yeah, now we are Oh, the validation was done. Wow. Okay. Now we're now we're really blitzing through things. Let's go. All right. Um, let's see. Wow. All Wow. Okay. All tech checks passed. Nice. So, where are we at now then? Let's take a look. Um, oh, it's deploying to Verscell already. It's doing this off uh camera here on my left monitor. authorizing Burcell. It's It's doing everything. It's ripping through it because we're on the Yeah, we're on the deploy now. We're Oh, we're almost done already. This is awesome. So, the integration took a long time because there was so much work to do, but then it like nailed it. And so, now we're just ripping through this workflow. That's exciting. Cool. So, it won't it won't have to be a three-hour live stream after all. All right. Nice. Let's see. not a good time for teaching in general. Almost no one cares about buying tutorials and courses or any kind of learning material. I agree to an extent like pe people are hungry for knowledge but like attention spans aren't the best right now and everyone just wants to like talk and get answers really fast from large language models even though they're they're oftentimes not the best answers as far as like higher level learning goes. Um, so yeah, like I I think like one thing that is a big bummer is um like longer coding tutorials are like totally dead [snorts] and like I feel like my channel has always been AI focused, but I've always sort of been in the middle where I'm not just like here's the latest thing with Claude. I'm also not doing like a you know twohour JavaScript application build tutorial like JS Mastery or like uh Travis or some someone like that. But I'm I'm I'm in the middle. Like I'm still doing like longer in-depth tutorials like using coding agents, but like doing a live build. I would kind of call it like a a deeper teaching video and like yeah, those unfortunately don't perform as well anymore. Like people just want the the hype stuff fast a lot. And uh it is kind of unfortunate cuz I used to be able to put out videos like I remember one video in particular that I made. It was uh a tutorial on Langraph that I that I made I think at like the very end of 2024. And that video ended up being 50 minutes long. Here I can even pop it up right now. This is like the perfect example of a video where it was like really really in-depth. So much so that like I actually thought it would probably perform poorly because of like how much like detail was there. See if I can find it. But the video like actually took off. It did super well. Um, how far back do I have to go? Yeah, I have I have to go really far. Um, yeah, this one right here. Uh, so this was like I actually I really really loved making this video combining pyantic AI and langraph which by the way I have a whole module in the Dynamus community in the AI agent mastery course for combining the pyantic AI and langraph are still fantastic. Like obviously I'm doing a lot more like AI coding content now, but as far as like building production grade AI agents and orchestrating them with Langraph, like this is still the stack in my mind. But yeah, this video 171,000 views, but it's a 51 minute video and it's very in-depth. Like I'm going through I'm showing the code, all the code that I wrote. And at this point, I was I [clears throat] still wrote all this by hand. Like literally, I did not use an AI coding assistance. I am here in VS Code. No cloud code. It didn't have wind surf at this point. Yeah, it like this was fun and but yeah, it was so in-depth like I this kind of video would not perform well these days. I mean, not just because like people use AI coding assistance, but just like in general like it seems like people don't appreciate this kind of depth anymore, which is a bummer because like I like I like doing that kind of depth. I know a lot of people don't and they're excited that like no one wants these longer tutorials, but I like making these longer tutorials. So anyway, I wouldn't call that a rant. I was I was going to say like thanks for for being there for my rant. I wouldn't call it a rant, but it's like appreciate you guys going down memory lane with me. Yeah, there there are longer YouTube videos that perform well, but they're usually like those crazy 4hour videos where it's like a novelty item and people like it more just because it's like fascinating versus I don't know. I mean, I'm sure a lot of people like actually go through the full thing. Lack of attention, squirrel. Exactly. Yep. Yep. All right. All right. So every node ran across provider blah blah blah. So yeah, the integration ran for 31 minutes but everything else was actually very very fast. So we got a real uh full stack app. Clerk O is working. Deployment is awaiting our input. So it ran clerk deploy um authenticated with verscell confirmed. There's no link yet. So it it can't so cloud code can't drive the interactive wizard that clerk deploy is. So I think I have to run this myself. Okay. So we have the artifact here. It's saying that I need to go through all this. I might do that in a little bit. Hold on. Let me let me talk to it here. So give me the path to the work tree like exactly where I need to run all the commands that you listed here in steps one through six. And then I also uh just so we can test things locally here. I want you to start the application. Give me the URL and the port so that I can uh visit it myself and test it all myself locally. So, we're asking for two things in one here because I I do want to I'm not sure if we we'll be able to get through everything here right now, but I want to run the deploy and like go through this wizard so we can get everything deployed to Verscell. That would be so cool because that would mean that this entire workflow took us from a spec all the way to a production deployed application with full authentication and apparently a Postgress uh database as well. I didn't even know that was in the spec, but that's I mean, it makes sense. It might be for Archon. I don't know. I don't know if it's going to try to deploy Archon or what, but we'll we'll find out. All right. Okay. Launching the server. It's saying that this is where I need to open up Claude as well. Okay. Yeah, take a look at this. We got a full codebase here. Let's take a look at the readme. Um, okay. So, it didn't it didn't actually create the readme. it. This is just the next JS starter read me, but we have uh everything here. Let's take a look at source apps dashboard. Nice. Where's the clerk off? Oh, yeah. So, we have like our signin page here. Nice. [laughter] This is cool. I actually don't know. Okay. Cool. It's running now in local host port 3000. Let's Let's check this out. The UI The UI looks okay. I'm actually not very impressed with the UI. Maybe once we sign in the the dashboard will look better. There's not really that much substance here. I guess it can't really be the best. the UI is okay and and like I said it might be kind of like something to adjust in the workflow where we can't we don't want opus in the planning to like dictate exactly what the UI looks like up front. So anyway, let let me sign in here. Um continue. Couldn't find your account. Okay, let me continue with Google. All right, sign in to clerk. Boom. Signing back in. Okay. Now, this is working well. All right. Uh Okay. Nice. Okay. This actually looks pretty good. Like the UI is kind of basic, but it looks clean, I guess. I I I guess I wouldn't say it doesn't look kind of LLM generated. I don't know. Maybe Opus could have could have done this. I probably should have been a bit more specific about how much I wanted in the UI to so that we could like really see something more like this other example that I had here where um it like looks really fancy like this is this is what Gemini 3.5 flash can do but it it still looks pretty good. So anyway, I don't even know how to work this application. We just built it built it. Um so pick a workflow set config A versus config B. Uh, tell the dashboard where your archon workflow folder lives before we can list anything. Oh, this is interesting. Wait, what? How do I I can't really see. This is like kind of jank. It messed up the UI a little bit. Um, I can't really tell. Where did I actually go for that? Um, tell the dashboard where your Archon workflows live before we can list anything. Open settings. And there's no Oh. Oh, here it is. Okay. So, I have to tell the dashboard. Okay. Let's do it here. Copy path. Oops. Okay. Save. Six. Workflows found. Workflow. Okay. I mean, this is working. This actually looks pretty neat. Like it's pretty clean. Um, so if I want want to run like uh piv system evolution. Let's benchmark this. How does this work? Okay, this is actually very cool. This is working. exactly as I hope. So the way that this works is you pick an archon workflow and then you describe the task. This is exactly what I want to benchmark here. And then you build the matchup. So like for each node, you get to define the model that's being used. And so I can say like let's do a test where it's um sonnet for everything. And then let's do a test right here where it's like um I don't know, let's use codeex here. Let's use we can like compare. So, it's the same workflow, but we get to use different models. Like, this is actually pretty cool. Let's do let's do codecs for everything here. Um, geez, there's a lot. This is a big workflow. Oh, yeah. This is the massive one that I built in the last live stream, I think. Um, okay. It's kind of getting cut off here. Okay. Yeah, the UI the UI has some work to be done for sure. But the the other thing is I didn't really allow it to iterate. Like part of that is on me because I didn't really give it any skills like agent browser for browser automation. I mean there's so many ways that I could have made this workflow better. I could have split up the plan more like I was talking about earlier. I could have given more steps for validation. Um so I yeah I don't really like blame any of the imperfections here. I just wanted to try to like show an example of mixing providers. Um, so yeah. Okay. Provide a prompt message before launching this matchup. Oh, yeah. So, I have to describe the task you want the workflow to do. Both A and B will see the exact same message. Um, build a landing page for Archon. I don't know. Okay. So, I can click start benchmark. Is this actually going to run? Uh, benchmark cued. Watch it run. Is this actually going to work? I I doubt this is going to work first try. per node breakdown. It said they were running, but it it looked like it crashed. Let's take a look at the logs. Um, yeah, I don't know if this is actually working. I mean, I definitely don't expect this to work. The point of this live stream was not to get a fully working application because this is insanely complicated. The way it has to integrate with Archon is insanely complicated. I more just wanted to see what it look like to combine providers to build some kind of front end. Like honestly, I maybe should have chosen something simpler. But still, this still is really cool. like it it's it's genuinely working at least up until the workflow execution because now I don't really get to see any logs. I think that um and it looks like Archon is down right now. Hold on. Let me see. Uh looks like Archon crashed. I need you to restart it here. Some something with my coding agent must have ended this session somehow. But it's very cool how I was able to import my workflows and then benchmark here. Like I can pick the different What? Wait, what is big pickle? What is this? Why is the default model big? Is that actually a thing? Does open code have something called big pickle? That's actually a thing. Big pickle is legitimately a thing. That is crazy. Well, anyway, it's it's very cool that you you can actually like set like this is working. It parsed the workflow and it knows each one of the nodes and it allows me to pick the model for every step. Like this is pretty good. I just don't know if the like the actual run is working. Um, oh, I see what's happening. Archon and this application are taking the same port, so that's why things aren't working. Okay, I got it. I got it. So now, now that I restarted Archon, it killed the port here, right? If I go to dashboard 3000, then Yeah. Okay, that makes sense. So I I kind of killed it again. Okay. Uh you need to start it on a separate port here. All right. Yeah, see it's on a 404 now. All right. I I'll let it restart. All right. So, anyway, while this is running here, it told me uh okay, all the commands that I have to run here. So, I'm going to pull this up. Let's open up a terminal here. Uh, we can just do this like side by side. Uh, actually, sorry, this this is going to be kind of hard. I want to full screen it. It's kind of hard to read everything there. Okay. G. Okay, there we go. All right. Uh, what are the commands I need to run? I kind of lost where I was now. Um, wait. Okay. Now, what commands do I have to run in the work tree? Because like Claude does this thing a lot where like when you switch the the tab between monitors, it like shves everything to one side. It's kind of jank. So now now I like totally lost where I am. Yeah, I'll just have it uh list them out again here. Okay. Uh oh wow, there's a lot I have to do still. Run these in a fresh PowerShell terminal. Yep, because they're interactive wizards that need real standard in. Okay. So, npx link versel. I don't know if I'm going to be able to like go through this end to end right now, but I'm just going to try so we can at least see the process here. But this really isn't the point of the live stream here, so I'm not going to spend an insane amount of time with it. Um, let's see. Okay, linked. Apparently, that's already done. And then I can uh pull I hope this doesn't display it. I mean I'll just delete it if it does. Okay. And then push the drizzle schema to prod. Okay. Output this to MD. I needed to It has multi-line stuff. So I have to have it output to a markdown quick. So okay. What what all am I doing here? So I I run the clerk deploy promote the clerk keys and then Okay. Yeah. So this is actually pretty straightforward. This is nice. Uh wow. So I I don't think we're going to get to the point where the deployed application can actually run Archon workflows. That'd be super involved to get all that deployed, but we're going to get the application deployed. So, from a standpoint of like planning it out with Opus, building it with Gemini, having something that is shipped, we're going and with authentication, like we're gonna have all of that uh which is pretty cool. So, okay, let me open this file. Deploy.m MD. Okay. Yeah, let me get this ready. Deploy.m MD. Where'd it put it? There it is. Okay, let's see. I did the linking already. clear. Now I just have to Okay, do that and then see if that works. Okay. Okay, now I do MPX clerk deploy. So this is the interactive one. Yeah. Okay. So let's see. Yep, that's good. Production domain example.com. Wait, I don't know. I will need a domain that I own, the ability to add DNS records on that domain. Uh, okay, sure. I guess I'll just do like um geminitest.dynamus.ai. Okay. [laughter] Or wait. Oh man, I don't Do I have a I have a separate domain that I Sorry, I'm trying to figure out do I need to use a subdomain? Oh, I think I should just use my root domain. Ah, screw it. I'll just do it. I don't care. Let's do it for the stream. I just I just don't want to like screw up anything that I have with my setup already. Uh, all right. I'll just do this. Clerk.dynamus accounts. Uh, let's do it. Let's do it. All right. So, I have to add these records. Uh, after OAS setup, you can verify DNS or skip export DNx as a bind zone file. Uh, no. Okay. I guess I'll do this on camera right now. I mean, it'll be quick. Let's pull up uh NameCheep here. Hold on. I'll do this off camera quick. So, I have to get to my account, my domains. Sign in. I I want to go through the full process here. Like, I feel like that it's just good to do that. Um well, actually, the only other thing is I could use the archon.diy domain or Yeah, that might actually be better. Let's actually do that. Let's use the I'm going to redo this and use the archon.diy. Oh, shoot. Wait. No. How do I go back? Maybe I can't. Oh crap. Wait. I'm kind of confused now. Mpx clerk deploy because I wanted to I want to restart. Oh shoot. How do I restart now? There's no arrow keys for that. Ah crap. Because I also I lost the record. I can't even go back to the other domain because I lost the records. I cleared the terminal. Uh is there like a clerk folder I can delete or something? Ah crap. Okay, I'm gonna have to go back and ask it here. Okay, I went partially through the uh clerk wizard and now I want to go back. But how do I do that? Like I want to set up a different domain. I need you to search the web. Shoot, I changed my mind and now I can't go back. I mean, I probably can, but I just don't know how. All right, the clerk deploys like brand new, so I'm not sure if a web search will help. Might just have to look at the CLI locally and figure that out. All right. Oh, big bat crack. All right. Well, I'll go to the chat while we wait for this here. All right. Move that over. Nicely done, dude. Gave me great inspiration, ideas for my work. Well, I appreciate that. Glad to hear it. Always nice to see your videos. Always learning something new from you. Thank you very much. All right. Uh, yeah. create an open router API key with a 1 cent limit. Yep. Um Sean already decided not to renew my enthropic subscription and switch to OpenAI. Yeah, I don't blame you. You know, the anthropic subscription, it has been better recently. Like my rate limits haven't been as bad, but there were a couple of weeks where I was like, "Oh boy, this is rough." Yeah. Is AI truly sustainable after the venture capital funding dries up? Some companies are proving that it it might be. Um, like Anthropic is predicted to be profitable a lot sooner than OpenAI. Now, it is kind of crazy how bloated their valuation is cuz they're they're valued at what was it $695 billion now, but then their actual revenue I think is like 60 billion a year. That's like way different than other companies. I I I saw some comparison where it's like Anthropic is valued a as much as um I I'm already forgetting the company. It might have been like SpaceX, but they have like way less revenue. So like other companies with the same valuation have like way more revenue. So like it is kind of crazy, but like they are predicted to be profitable. I think it was like 2028, which I mean they're funding they just did their series H, didn't they? Like they're they have the funding they need. Uh right wasn't it um anthropic series H they raised uh yeah 65 billion oh no sorry not not their valuation is not 695 it's 965 billion is there so they they're almost a1 trillion company now uh but their revenue is like pretty low for that valuation compared to other companies it's pretty interesting uh yeah their revenue crossed 47 billion earlier this month which is yeah very low like um yeah like uh what other company besides Anthropic has a valuation close to a trillion I'm trying to remember the exact it might have been SpaceX or was it Nvidia SpaceX value between 1.5 and 1.8 preparing for IPO. Uh, what is SpaceX's revenue? Wait, is it actually less than Enthropic? These are not the figures that I remember seeing. I'm sorry, guys. I don't I don't remember the exact company or comparison that I saw. I'm I'm trying to remember and it's uh obviously not coming together here. If if someone knows the figure that I'm talking about, please share that in the chat. There was something where it's like Anthropic is valued at 965 billion. Here's another company also valued at that much, but the revenue is way higher. So like that kind of like speaks to the AI bubble, but this this figure is not uh arguing that. Yeah, I don't I don't know the company. Maybe maybe it was Birkshshire Hathaway honestly. What is Birkshshire Hathaway's valuation and uh annual revenue? Yeah. Okay. This was the comparison. So Birkshshire is a is a little over a trillion like very close to anthropic but instead of 47 billion their their annual revenue is almost 400 billion. Like that that is crazy that crazy difference. Like that just shows how much the multiplier on your revenue uh is for your valuation for AI companies. Like it's insane. Like the 10x more or like the it's it's like a 20x, right? Like we take 965 divided by 47, it's literally like over a 20x valuation. Like a ar multiplied by 20 is their valuation right now. Like that's bonkers. Samsung is the one that you saw or Walmart. Okay, let's take a look. Um, what is the valuation of Samsung and their annual revenue? Um, oh yeah, that's very similar. So, yeah, about the same as Enthropic and Birkshshire Hathway. And then it's not as crazy as Birkshshire Hathway, but yeah, still like 235 billion is insane compared to Anthropics 47 billion. Like that that multiplier still that's a massive difference, but yeah, that's another And then Walmart here. Uh what is the ARR for Walmart and their current valuation? Walmart's probably nuts. Yeah, 700. They are uh holy cow almost a trillion and then what's okay what's their crazy so their valuation is only a little bit over a trillion or I guess briefly surpassed so maybe it's even less now but then their revenue is 700 billion like that is nuts Wow. All right. Interesting. Very interesting. [laughter] Well, anyway, let's let's see what we got back from. The wizard already created a production instance pointing Oh, nah. All right. So, I can't it's kind of too late probably. Yeah, because it it it already moved forward, which is cool. It created the production instance and clerk already. But I can I can change the the domain if I want. Yeah, the only the only problem is um so okay, I'll move forward with Dynamus.AI, but the problem is I forgot the records to set. Okay, what are the records I need to set because the CLI printed it out to me, but then I cleared my terminal. So well, I have it printed out and then I'll move forward with that. Did I use any llinters in this build? Are people generally skipping llinters? Uh, no. Don't skip llinters. I did. I did the linting here. If we look at the validation summary, it uh it used eslint. So, it did check that. Don't skip llinters. They're actually pretty important. All right, let's see. I use Hermes with codeex. All my archon workflows are codecs. Very cool. Sean Hermes uses the archon skill and I code on three prod code bases using Discord on my phone. That is that's awesome. Never got rate limited. Yeah, codeex might be the move. I'm definitely going to be exploring it more, especially with that um June 15th cuto off for anthropic going off of the subscription for the claude agent SDK. Yeah. All right. So, anyway, let me go back here. Um, here are the records I need to set. Okay. Okay. I I'll I'll try to rip through this really quick here. So, let me go back. I'll do this off camera just really fast here. So, domain list Dynamis manage my advanced DNS. Add a record. CNAME host is clerk points to All right, that copy didn't work. Okay, check. And then add another C name. All right. Boom. Check. Add another C name. [laughter] We'll get through this as fast as we can. All right, [sighs] that didn't work. Hold on. Slowest part of the stream. There we go. Check two more. CNAME. All right. >> [snorts] >> Oh, this is the same. Wait, no. No, it's not the same. Okay, these two are the same. Oh, no, they're not. DKM 2. Yeah. Yeah. Okay. All right. Copy this. Boom. All right. Okay. Done. Got them all set. So here here we go. Check marks across the board. So I wasn't showing that on camera just because I don't know if they're like I don't know if I want to just like show all my DNS records. But anyway, looking good. So let's go back here. Okay, I already have my client ID and client secret. Let's see. Shoot. I guess I'll have to I'll have to I thought I already had this though. Hold on. Doesn't deploy pick us export configure oath wait for verification because I thought I had these and environment variables already. I guess not. So I already have my client. I mean I do. Okay. So this part I definitely have to do off camera obviously. So let me do that. Okay. So I have to remember where I actually have my keys scripts. Where's my Google credentials? Okay, here we go. Client ID. All right. Then client secret. Oh, shoot. Hold on. The input is masked. So, I thought that I didn't paste it, but I think I did actually. I need to copy that again. Yeah. Okay. Okay. No. Okay. So, let me bring it back. Uh, verify DNS records. It might take a while. Like this might not work yet because it has to propagate. We'll see. [clears throat] I I want to get to the point where we have everything deployed. That would be so incredibly cool. [sighs] But yeah, this this might not uh be ready yet, but I think we can move forward with the rest of the commands before this and then we would just be able to check again. Let's see. Uh retrying in a minute. So, we got our exponential back off here. But yeah, we probably won't be good yet. But anyway, so after we go through this, then we just have to promote our keys. Archon stubs. Yeah. So, yeah, we're not actually going to be deploying Archon here, but I just want to get like the application up so I can like log in and really have it in production. And then going back to my second brain, it I think it gave me like the URL so I can actually see things in Clerk, too. Uh, yeah, here we go. Let me open this up. And I'll do that here. I'm going to bring this off camera quick just in case it is actually anything sensitive it's displaying. But it'd be nice to have a look at the clerk dashboard to see what that looks like because it's cool that we can literally manage everything from cloud code, but I still want to have like that visibility. Um, okay. Okay. So, it's it says that we're awaiting our deployment still. So if I go to configuration and I look at domains, it it should. Yeah. So unverified still. I set that on the right domain, didn't I? Wait, I might have to like check that quick. Did I? Yeah, I did. Okay, we're good. It's just not propagating yet. Okay, what if I verify here? Configuration verification successful. Why is it successful in it? Oh, you know what? I think I did it wrong. Yeah, I did it. Shoot. because you're not I did like you know like clerk.dynamis.ai but it's supposed to be just clerk. I need to actually edit each of these but I can do this like really fast. That's my bad. Okay. Check. Check. Check accounts dot um clerk.services. Okay, that actually should be good now. So, I'll have to wait a probably a little bit, but let's go back and see where we're at with the deployment now. Um, let me check again. [snorts] All right, let's see. Sean say I switched to Codex when they clamp down. $100 plan goes further than the Claw 200. I believe it. Yep. Quality is just as good. I think that one's a bit up for debate. Um, not I'm not saying I disagree, but I just like I I feel like a lot of people still think Claude is a bit better, but GBT 5.5 it's pretty good. Like I'm definitely not going to miss Opus if I have to switch my second brain to codeex. And like I said, I'll I'll have to see if that is the case or not next month. >> [clears throat] >> Okay, this DN I might do the DNS later. We'll see. Um, I'm confused. Why don't you make an agentic workflow that doesn't involve SDK or -p? That's how I set up my workflow. So, I'll be always just be using my subscription but running 16 agents. Well, the reason I'm using the cloud agent SDK is because I want it running in the background, right? So, like I can talk like I use the cloud agent SDK so I can talk to my second brain in Slack. So, it's running as a an application. I need to programmatically access Claude. They're they're really going to be um disallowing any use of your subscription with Claude unless you are in the terminal talking to it. Like interactive Claude is the only way you can use your subscription starting June 15th. And so what you're defining here like when you talk about running your workflows in the in the terminal, that's that's fine. But the problem is you're still manually invoking each one of the steps. I mean feel free to elaborate more if you'd like but like with archon we use the cloud Asian SDK because we have the application orchestrating the workflow for us handling the work trees handling the context switch between nodes and um everything that we have for like parallel execution like we have to invoke claude programmatically. I can't have Archon like be this feature that lives in my terminal if that makes sense. Like it just won't be possible. Okay, so I skipped the DNS. Okay, I don't know if this is going to work because I this isn't the obviously the domain where I actually deploy things. H we'll see. We'll see. Okay. So, I'm going to clear. I'm going to go back to deploy and we'll run each of these commands. Let me Oh, wow. It's making me actually uh add all these publishable key here. I I need I need to list these all out for me. Okay. For all of these values that I have to deploy as environments to Verscell, I need you to list them all out for me here. And uh then this this I have to bring off camera obviously. So, I have to enter these in. Okay. Um, and then let's let's try the verification one more time for the domains here. I swear I did it right this time. Now, I don't know why this part I might have to Oh, no. We're good. We're good. Okay. Nice. All right. So, we got we got the domain set and now it's just generating the asserts for us. Okay. Nice. All right. Wait, what? It just went back to unverified. Um, what the heck? It was just working. Uh, that's weird. I think it was just a bit of a glitch because we're good now. Okay, I'm just going to call it good. All right, so back here to deploy after we run each of these. Uh, yeah. Yeah, then we then just do this as empty values. Okay, actually I can do this already. I can do this right now. So, uh, let's run that. [snorts] So, archon home, this environment variable doesn't have to be set. No, it's not sensitive. I'm going to set it as empty. Uh, oh, it doesn't let me set an empty value. I'll just do that. All right. And then do the same thing for this one here. My off camerara is confusing me here. Hold on. I got to take a look at what it's saying. Um, let's see. It didn't give me the actual values here. I want you to list out all the values for me here, like the actual ones, so I can copy and paste them. Okay. [snorts] And then what else can we do here? Can add the public app URL. Um this is not sensitive. We'll do um geminitest.dynamis.ai. All right. And then we can ship it. Oh, I can ver Can I verify? Hold on. I'm going to test this. Checking for production instance. So, we're making sure authentication is set. Waiting for clerk DNS check to process. Oh gosh. So, the DNS records are still not propagating yet. So, we might not be at the point where we can. So I might not be able to do this right now because uh we have to wait for the DNS to propagate. So I can't really do the full deployment right now. Yeah. But at least we got to see the process. It it it's super cool how easy it is to get the authentication wrapped up in the application. I we saw the whole thing work live. Um let me actually have it restart the app again. Wait, what port is the app currently running on? Otherwise, uh go ahead and restart it. So, we'll we'll uh take a look at it one more time here. But we have the deployment is pending, but the local setup is fully working and we have the authentication with clerk. It was so easy to get that set up. We have it here in the dashboard. Let's see if I refresh how we're looking. Um, one of these is verified. Okay. Yeah. So, it's just taking its sweet time because of the DNS propagation, but we'll get there. Uh, the app is running on localhost 3001. Okay, let's take a look. So, if I sign back in, yeah, I mean, it's pretty pretty sweet how easy the authentication was. Piv system evolution had a timeout. Um, that's just because we had our archon shut down. But uh yeah, I mean this is looking pretty good. I would say that like I would definitely want to work on the UI more. I still feel like this looks better than what Opus would have done by itself. I I I I like the feel of it. Like it actually looks pretty clean. I mean things like this little blue bar right here. Like LLM's always love to do that. So there are definitely some LLM isms right here. Still looks pretty good though. I don't know. I I have to do a lot more experimentation with this. But really like what I wanted to do with you guys in this live stream is just start that exploration of like what does it look like to have a workflow handle a larger task using Gemini for the UI design and claude for everything else. And like we tested that successfully. I mean we don't have like the most conclusive results right here but building live in a live stream we can't always expect to like we we just went through the process together. It was fun to build that workflow and we went through the full workflow. We have something that I mean it's almost there to deployment. It's pretty cool. like we only have to do a couple of more things and then we'll get there. So, I'm going to keep working on this over the weekend. I want to get this like fully deployed and everything. Um cuz this is also like an application that I'm genuinely interested in building for myself and then like maybe building some kind of benchmarking thing into Archon. This is also kind of like the initial fora into that as well. U because I want to make it possible for people when they're building their harnesses in Archon to be able to really evaluate like how good is this harness? like how much is it actually solving the problem that I have for it? Like the kinds of tasks that I wanted to go through with it. All right, cool. So, yeah, I'll I'll uh chat with you guys here and then maybe we'll come back to this. Otherwise, I might uh end the stream um if we're still waiting for these things. But uh yeah, I just want to chat with you guys for a little bit to end things off here. All right. Um, would it make sense to have some kind of hook implemented to prevent port collisions? Uh, definitely that is a good idea. Yeah, I I've actually not really ever had that happen before. Usually Claude like knows when something's running on a port or like it will look at that and then start it on a different port. Uh, but yeah, that's that would be a good hook to have. That's one of the benefit or one of the really good examples of hooks is just to like make sure that uh your environment is isolated. You're not overriding ports or editing files in different directories and things like that. All right. Uh great live stream as always. Thank you very much. I've been watching all them. Wanted to ask, have you ever considered using local LLMs for sub aent work? Do you think that could be useful in practice? Uh definitely. Yeah. So I I always use sub agents for either web research or codebased exploration because that's where just having some summary returned to your main agent is all the information you really need. So not not sub aents for implementation but just research and and local models are certainly good enough for those simpler tasks. any kind of research task and definitely like at least like a 20 billion parameter model like one of the newer Quens could be good or like Gemma 4. Um certainly could be good for sub agents. Yeah. So I haven't done that much exploration with local models recently and honestly the biggest reason is I have this weird issue on my computer. I think it's something stupid with WSL um where like my computer will crash when I run local models. it. It's so unfortunate because I I literally built this computer to run local models and I've been having a lot of problems with that recently. So, I need to figure out I've been continuing to to figure that out. Like, yeah, it's it's annoying trying to fix my computer actively right now. There's a lot of things I'm trying with driver updates and it's yeah, it's kind of a mess, but yeah. Yes, I'm very interested in doing more testing with local models soon I figure that out. Um, yeah, retail companies get low multipliers. That makes sense. Birkshshire Hathways, it's an investment company. Yeah, like I totally understand different industries, you have different multipliers, but but like still the multiplier for AI companies is bonkers. Like anthropic is just bonkers. Workflows are pointless for 99% of people, right? I've heard they destroy usage limits like crazy. Are you talking about the new dynamic workflows in Claude? Because that might be true, but also for any kind of sub agent, you can choose to use a different model like Haiku or Sonnet, like if you don't need as much reasoning capability. And then like just in general when like if you're talking about more workflows in general, like archon workflows for example, it's the same thing where it's like each node, each step of the workflow, you get to pick the model. So based on the reasoning power you need there like if you're just doing exploration then don't use opus just use Kimmy K 2.6 or a local model or use haiku or sonnet instead if you want to use claw code for everything. you always have the opportunity to uh make your workflow more token efficient because like I proved out in recent live streams, you really only need the most reasoning power in certain steps, usually the planning, and then the rest of the process can be pretty much as good with smaller models because you have that better spec that's guiding the entire implementation and even all the validation that you do. June 16th, Mythos will be available. Yeah, that'll be the day. It's kind of crazy that like at first Anthropic is like Mythos were keeping it only for the big companies because it's a big security risk and then like all of a sudden they're just releasing it to everyone. It's It's weird. Yeah. Anyone else loving the feel of Opus 4.8? I'm liking it. It doesn't feel like significantly better than uh 4.7 to me, but one of the things that they've that Enthropic advertised, which people seem to say like this is this really does hold up is that it just hallucinates less in general. Like Opus 4.8 for the first time is a model that can like really reliably admit when it doesn't actually know something instead of just fabricating information. So it seems like hallucinations are down quite a bit with Opus. So, it's not necessarily that like the code quality is way better than Opus 4.7 like raw output quality, but there's just less hallucinations, which I mean that I guess that does lead to better quality overall. So, it is definitely a pretty decent step up. Um, wish that the Claude app could have a sidebar like CodeC. Yeah, that would be nice. The Codeex app is is really good. Yeah. Uh, am I anticipating a faster or slower AI field evolution? Which direction is my preference? Well, it kind of depends on what you're talking about. Like there's different uh aspects to the AI field. Like you have like regulation, you have the power of the models themselves. Like obviously I want the power of the models to to go up and up and up and like we're kind of at a plateau right now which is sort of unfortunate as far as like regulation though. Like I think one thing that does need to slow down is uh how much companies are just like blowing through money and trying to just like it feel feels like companies don't really have the greater good in mind right now just in general like slow down and like think about like how is generative AI really going to have an impact on society and like what are guard rules that we need u because I think there are some like real concerns right now for job displacement and like a lot of it is very overexaggerated and I'm not like I'm not like a super doom and gloom guy for this but I do think that like like just the greed in the world in general is like kind of unchecked right now. I guess I just put it that way. I mean it's something that like I don't think we need to dive super deep into right now. But I do think that like the world needs to slow down a bit together and like think about longer term effects of some of these things. Uh, but just because of capitalism, market pressure, comp competition, whatever you want to call it, it's no one's slowing down. And I guess that's a little contradictory where I just said like I'd want the power of the LLMs to go up and up. Um, so yeah, I guess you can't really have both. But yeah, let's see. Let's see. Like the little glow in the top right. Um, looks nice, but definitely an LLM ism. which glow. I I thought I remember seeing something like that. I don't remember where that was now. But yeah, there's there's there's a couple things here that are LLM isms, but it still is like like it looks it's a pretty good application. Like it looks really clean for a dashboard. I'm your big fan. I appreciate very much. Thank you. My WSL2 works well for local models. So might be your config or something. Yeah. Yeah, I mean I got to keep diving into it. It's very annoying though cuz if I like if I were to try loading a model into VRAM right now, I the stream would probably end. It would just crash. Like it's unfortunate. All right. [snorts] I enjoy live, but I'll start watching hour late at 2x. Fair enough, Jeff. Yeah, I appreciate you being here, though. All right. Um, let's see. Oh. Oh, you were replying to someone named Jojoy. I see. Sorry. I I just could realize that. Amazing as always. Just started watching in 1.2 speed to catch up. Yeah, sounds good. Cool. All right. I agree. Less hallucinations is far a better implementation plans. Also, pause to do far more Q&A. Grill me without asking it to do so. So, yeah, it's right. It kind of like helps you steer the process more, too. like it kind of pairs more with you. Yeah. Like it's more of like a Opus 4.8 feels more like it wants to like work with you, I guess. Like and and like not just kind of like run off with things and fabricate information. It is nice. Yeah. Cool. All right. Well, anyway, uh I think let's let's check the clerk config here. It's uh yeah, still unverified. So, we'll have to wait for the DNS propagation. So I won't be able to do I won't be able to fully finish the deployment here, but we got to the point like we ran every step of the clerk authentication. We have that integrated in our local environment here. So yeah, I mean I would say this was a success. Like I definitely have to do more experimenting in general with Gemini for building and like Opus for planning. I want to try tweaking it more so like Opus doesn't try to dictate the layout as much in the plan and see if that helps Gemini run with it more creatively as well. So I'm going to keep doing some testing. I'm also interested in uh putting out a video that kind of like compacts a lot of my testing and learning here on my channel. So, I might be doing that uh either Monday or Wednesday next week. Um so, yeah, I'm going to keep working with this. So, like this live stream is all about like let's build it live. Let's kind of get into this and see what it looks like to mix providers in this way and then uh yeah, I'll keep polishing it up. So, that is the plan for this week here. So yeah, I'm going to go ahead and uh call the stream here. But uh I appreciate all you guys being here. This this was fun. This is a long live stream. Like holy cow, time flew by. This is two and a half hours. But yeah, it was fun. So yeah, appreciate all you guys being here. All your questions and comments and engagement. It was funny that the stream started so slow because I accidentally scheduled the stream for 900 p.m. instead of 9:00 a.m. That was crazy. Uh gez. But but but we you guys showed up and this was fun. So all right. I hope you guys have a fantastic rest of your weekend. Stay tuned for uh the video here where I'll kind of present some more polished findings for everything that I'm building here. I'll share like a final workflow and stuff if you guys want to use it for yourself as well. So, it'll be good. But yeah, I'm going to go ahead and sign off now. So, yeah, have a great rest of your weekend, guys, and I will see you around. Take care, everyone.