19:10Pat Simmons · Tutorial
Clone a $1.2B App in 19 Minutes
An 8-step agentic pipeline that takes you from naive AI slop to a pixel-near Linear replica, deployed to Vercel with an MCP server, in under 20 minutes.
June 8thThe argument in one line.
A single AI agent-management app can be built end-to-end by iteratively critiquing an AI builder's output rather than writing UI code, using existing open-source projects as both design reference and a source of reusable architecture.
Read if. Skip if.
READ IF YOU ARE…
- You run multiple AI coding agents across different tools (Claude Code, Codex, Telegram/Slack bots) and have no system tying them together.
- You want a concrete example of iterating with an AI app-builder through multiple rounds of specific critique rather than a single big prompt.
- You're curious how to give an AI agent persistent memory/identity across sessions without building that system from scratch.
- You want to route open-source models (Kimi, MiniMax, GLM) through the same interface you already use for Claude, via OpenRouter.
SKIP IF…
- You only run a single agent or work exclusively in one CLI already, so the multi-agent dashboard problem doesn't apply to you.
- You're on Windows. The finished app only ships as a signed Mac DMG; Windows users are told to self-port.
The full version, fast.
The video argues that AI agent sprawl (different agents living in different apps, repos, and platforms with no shared system) is solved by building one custom control app rather than adopting an existing platform. The creator uses Claude Fable 5 as the builder: he sketches a rough mockup, references two open-source agentic dev environments (Superset and conductor.build) for layout, and iterates through blunt rounds of feedback (kill emojis, connect to real terminals, add multi-model support) until Fable produces a working Mac app. The key architectural move is wiring OpenRouter's Anthropic-compatible endpoint into the Claude Code harness so other models (Kimi K2, MiniMax, GLM-5.2) run inside the same UI, and borrowing Hermes' open-source agent-memory file structure (agents.md, user.md, memory.md) instead of designing persistent agent identity from scratch. The app is then published as an installable open-source repo.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:22
01 · The chaos problem
States the multi-agent sprawl problem and previews the harness he's about to build.
01:22 – 03:25
02 · The mockup and reference apps
Sketches the layout idea, then shows CMUX and conductor.build as existing ADEs to reference.
03:25 – 04:59
03 · First Fable build
Feeds Fable the prompt plus a Conductor screenshot; gets a first web mockup in about 20 minutes and gives blunt UI feedback.
04:59 – 06:48
04 · Naming it daemon
Names the app daemon mid-build and explains the Greek etymology while v3 finishes.
06:48 – 07:42
05 · Real Mac app
Tells Fable to turn the web mockup into a proper Mac application; about 90 minutes later it's a working native app with categories, agents, and repo cloning.
07:42 – 11:31
06 · First agent live
Sets up Ernest the script-writing agent, clones an empty repo, opens a live Claude Code terminal session already authenticated.
11:31 – 13:56
07 · Adding multi-model support
Asks Fable to add Codex/ChatGPT login and open-source models (Kimi K2, MiniMax, GLM-5.2) via an OpenRouter API key using the Anthropic-compatible base URL trick.
13:56 – 14:57
08 · Fixing the auth bug
Hits a missing-authentication-header error running Kimi via OpenRouter; sends it back to Fable for a quick fix along with a logo change.
14:57 – 16:57
09 · Borrowing Hermes' memory system
Realizes agents lose all context when a tab closes; tells Fable to port Hermes' agents.md/user.md/memory.md structure wholesale.
16:57 – 17:57
10 · Configuring agent identity
Opens the ported agent files, edits agents.md by hand to define Ernest's real role as a YouTube scriptwriter.
17:57 – 19:21
11 · Publishing and onboarding
Publishes the app as a public GitHub repo with a signed DMG, walks through install and first-run setup for viewers.
Lines worth screenshotting.
- Multi-agent chaos comes from tool fragmentation, not agent count. The fix is one shared interface, not fewer agents.
- Cloning an existing open-source UI's screenshot into an AI builder is a faster design brief than describing a layout in words alone.
- Pointing an app's Claude harness at OpenRouter's Anthropic-compatible base URL lets any OpenRouter model impersonate a Claude backend with zero custom integration per model.
- An agent that loses its identity and memory every time you close its tab isn't actually agentic. Persistent markdown identity files are the missing layer.
- Rather than inventing an agent memory system, reverse-engineering and porting an existing open-source implementation (Hermes) took a fraction of the design time.
- A three-to-four round feedback loop (sketch, critique, rebuild, critique again) was enough to take a rough mockup to a shippable native Mac app in about two hours of build time.
- Naming a background-process manager after daemon ties directly to its original Greek meaning, daimon or guardian spirit, a naming choice that doubles as a description of the software's job.
Borrow architecture, don't just borrow inspiration.
WHAT TO LEARN
The fastest path to a working custom tool is naming the exact open-source project that already solved your hardest sub-problem and porting its structure directly, instead of designing it from a blank page.
- Screenshotting a competing tool's UI into an AI builder is a faster, more precise design brief than describing a layout in prose alone.
- The hardest sub-problem in a multi-agent tool isn't the UI, it's giving each agent persistent identity and memory across sessions.
- When an existing open-source project has already solved your hardest sub-problem well, telling your AI builder to reverse-engineer and port that specific implementation beats inventing your own from scratch.
- Routing alternate AI models through a base-URL swap (an Anthropic-compatible endpoint via OpenRouter) can unlock multi-model support without writing a separate integration per provider.
- Blunt, specific feedback across a few rounds (name exactly what's wrong: emojis, light mode, fake terminal) converges faster than trying to write one perfect initial prompt.
- Shipping a working prototype as an installable public repo with clear onboarding steps turns a personal build into something an audience can actually reproduce.
Terms worth knowing.
- ADE (agentic development environment)
- An app layout pattern for running multiple AI coding agents side by side, typically with agents/repos on one side and session tabs on the other.
- Daemon
- A computer program that runs continuously in the background as an independent process, without direct user control.
- OpenRouter
- A service that provides a single API endpoint to access many different AI models (open-source and commercial) from multiple providers.
- Anthropic-compatible base URL
- An API endpoint shaped like Anthropic's own, letting other providers' models be swapped in behind tools built for Claude without code changes.
- AGENTS.md / memory.md
- Markdown files that define an AI agent's role, operating instructions, and accumulated memory, updated automatically as the agent works so it retains context across sessions.
Things they pointed at.
02:25toolCMUX
02:32toolconductor.build
02:58toolSuperset
15:00toolHermes (GitHub repo)
11:45toolOpenRouter
04:57linkPrompting Fable 5 (Anthropic documentation)
Lines you could clip.
00:00
“AI agents are everywhere.”
10:46
“This is really nice because you can bring in an existing GitHub repository, and it already opens up Claude.”
15:00
“I know Hermes has probably the best agent memory that they've configured, so I can give this to my agent and have it parse through how they've structured the memory.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00AI agents are everywhere. But there are two problems. First, pick one and you're stuck with it inside their model or inside their platform.
00:07Second, it is complete chaos. You might have a pile of chat windows inside different apps like Telegram and Slack, different repos inside Cloud Code and Codex, and absolutely no system to any of it. If you're using an agent platform, the context switching alone is killing your productivity.
00:21But what if I told you there's one app that will run all of them? Any agent, any model, all in a single place. That is exactly what we're gonna build today.
00:28And no, this isn't Hermes or OpenClaw. It runs on the subscriptions you already pay for. We're not going to get into any ToS gray areas.
00:35And all of your agents, all of your chat windows can be accessed inside this platform we're about to build. So in this video, I'll show you how to build out this agent harness completely from scratch with Fable five. We'll get it running in a single command, and by the end, you'll have the whole system ready to start building out your own team of agents even if you've never built an agent before.
00:51So not wasting any more time, let's dive right in to the build. But first, let me explain what I'm envisioning for this harness. So here is the mock up I sketched out.
00:59Please ignore the chicken scratch. But the basic idea is we have a category on the left for each one of these agents. A good category could be, say, YouTube.
01:09And under the category, we have different agents assigned to different parts of YouTube. The first agent could be a script writer agent. The second agent could be a motion graphic creator agent.
01:19Another agent might monitor analytics. And the big idea here is if I select an agent, once I select that agent under the category, then I can open tabs here, almost like Chrome browser tabs, and each one of these will be a terminal session.
01:34That will allow me to connect to different models, whether it's Opus, GPT, or any open source model. And each one of these terminal tabs could be three different scripts I'm working on at one time.
01:45I, for one, am incredibly scatterbrained, you might be too, and I want a system that can keep all of the different things I'm working on nice and organized. Because typically, I'm working on multiple scripts at a time, and the the same will probably be true for editing or motion graphics, for analytics as well. So that's the basic premise, and this is not a whole lot different than how they've architected codecs.
02:07If we look at codecs, we could, in theory, in these projects, organize it in a way where each project has a folder assigned to each agent. However, we still run into issues when I wanna create new tabs at the top.
02:19The tabs are in the left hand side here instead, and I could be chatting with different agents at one time. But if I have, say, I don't know, 15 agents, it starts to get really cumbersome when I have all these different projects I have to scroll through. If that's confusing at all, let me show you a few examples that get closer to what I'm thinking.
02:35They are called ADEs or agentic development environments. There are a bunch of them.
02:40CMUX is a good example. The way they have architected this is they have these repos or agent tabs on the left, and then they have tabs within each agent. However, these tabs are kind of all over the place.
02:51It it starts to get really confusing, but this is more like the general premise. There is another healthconductor dot build that to me is a little less confusing. Agents on the left, tabs at the top, but each of these still feel more like agentic coding environments.
03:05And a lot of these agents that I wanna build out aren't necessarily focused on coding. So we're gonna take pieces from each of these, and the beauty is they are open source. So I can go to their GitHub repos and start to together into something that I'm actually happy with.
03:20So that being said, let's now go back to Fable and start brainstorming this. So we will open up a new session in terminal here with Fable five model Fable.
03:32I'm gonna paste in this prompt. I will put it on screen. Essentially, what I am saying is I'm building an ADE.
03:39I run agents for everything, writing content coding. I want one app where all of them will live. I give some examples like the ones we discussed.
03:47I say start from superset actually because I take that back as I looked further into this. Conductor. Build is not open source.
03:54So what we're gonna do is we're going to use superset instead, which is very similar looking. Here is superset, same kind of idea. They have an open source repo I can pull from, and then I'm gonna take some screenshots from the actual Conductor interface, paste that into Fable as well.
04:07And then the rest of the prompt is basically just describing what I already went through. So the rails, the terminal tabs, each workspace has its own repo, and I talk about onboarding as well, which we'll get into once we actually get the the UI set up. But the basic idea of onboarding is I want it to be as simple as possible for the user.
04:22So the minute they open the application, they are asked to create agents, agent categories, and then agents under those categories. So that's the idea.
04:30And what we're doing as well is we're using a Fable orchestration skill that I put together based on this anthropic documentation here titled prompting fable five. But they get into a lot of the details around this. So I fed this into a separate agent, and we built out a skill based on this.
04:46And that skill can be found here. I will include this in the description below, but it is more or less what Anthropic Details and their docs here just organized into a nice easy use skill for an agent. And in about twenty minutes, we have our first mock up alive.
05:02Interestingly, it used clauder artifacts here. I'm not exactly sure why.
05:06Maybe that's a fable thing, but I mean it should work just as well as a local host. So if we go to this link here, and here it is. It's looking pretty good.
05:15I fed it a screenshot of Conductor, and you can see how similar it is to Conductor. We have categories on the left under those, the actual agents. These emojis, of course, are obnoxious.
05:26I don't think we need changes. It basically just matched exactly what I gave it in the screenshot. So it added changes, checks, run, terminal setup.
05:34We don't need all of that. A lot of this is superfluous. It also added a light mode, so that needs to be fixed.
05:40These emojis, of course, are driving me absolutely nuts. But I like the little tab here. I like the loading indicators.
05:46It's organized by agents. We can add a new category. We can name it, but create category, and that actually works nice.
05:52And then you can add an agent to test. Cool. And you can go different models, which we will get into setting up.
05:57So this is not a bad first pass. I'm gonna give it feedback. One big thing that I see is it's not actually showing terminal, which it needs to actually connect to a terminal window.
06:07As in, it needs to run directly through the CLI, and it looks like it's just like its own interface right now, especially when I go light mode. You can see it's, like, got its own little interface. So I'm gonna give Fable all of that feedback, and then we'll come back for version two.
06:19And just in case you're curious, here is the follow-up feedback that I gave. I will put this on screen now, but it's basically describing exactly what I said. Getting rid of emojis and ability to add a profile photo, connecting the CLIs, stripping back some of these extraneous designs and functions here.
06:34And while that builds, Claudine, I just thought of the name of this application, which is gonna daemon, d a m o n, which comes from a play off of daemon, d a e m o n, which if you're not familiar, is a computer program that runs in the background as an independent process. And actually, back in the day, they came up with this term based on the Greek for d a I m o n, which in Greek was a supernatural entity often seen as a guardian or or spirit.
07:00So it's kind of like a like a little helper, which is fitting for this agent application. Anyway, back to the build. And v three is now live, so let's check out this latest version.
07:10It is looking pretty good. Looks like most of my feedback is in. Agent files on the right.
07:16We've got these tabs at the top. We've got our categories. And then under that is agents.
07:21We have these nice little loading indicators. New category. Upload photo.
07:25Does this actually work? Just use this random photo from my Fable five video game test.
07:32Oh yeah. Look at that. Works.
07:33K. Cool. So we are all set up, ready to go.
07:36I'm gonna go back to this agent and tell it now to build this out into a proper Mac application. Okay. And in about, I'd say, an hour and a half of building, we have something ready to go, a Mac application of this agent's platform for us to react to.
07:52Fable is not fast, but it is incredibly thorough. It had spawned a ton of different agents and just went through this entire build process. So let's see what this app is looking like.
08:01It's telling me to run this install command. I'm just gonna say, can you run this yourself? Okay.
08:07So it looks like we have our Mac app up and running. We've migrated over from that mock up. So let's check how the Mac app is looking.
08:15How do I see this? And here it is. So it's looking pretty good.
08:20It's pretty bare bones, but we should be able to add teams here, what I was calling category before. Upload a photo. We're gonna do YouTube first.
08:29I have this image, YouTube. I'm gonna call this YouTube.
08:32Create a category. Then And we have an ability to under this, exactly like I described, looks like this is working. Adding a new agent.
08:40We'll call this it's a good script writer name. Earnest.
08:43And I can just add a photo here. Just gonna do Random photo of something. Here we go.
08:47This video game screenshot from my Unreal Engine MCP video. Good enough.
08:52And then right now, it's only working for Claude. So the runtime is gonna be Claude as in we're using our Claude subscription. And this, I actually didn't even ask for this.
09:00Maybe this came from Conductor or Superset, but this is really nice because you can, if you'd like, bring in an existing GitHub repository. Say you have an agent living in a GitHub repository. Maybe you already have agents and they're not in GitHub yet.
09:12You could theoretically put those in the cloud on GitHub, then upload them directly from here with this URL and just clone them and automatically upload them into our daemon app here. So I'm just gonna do new empty repo, and then I believe it's going to create a folder locally. Oh, wow.
09:28Okay. It already opens up Claude. So okay.
09:30Yeah. Here's the folder that it's opening up locally dot a d e default agents, and then it's just this string and then slash work tree. Then I can go, yes.
09:38I trust this folder. And boom. So we have Ernest, our script writing agent, all set up.
09:42You can also go to new tab, create a new session here, another session with Ernest. We should be able to, if I go command I, yep, change the name of this. So say this is a video about AI tools, and then and this other one could be about, say, codex.
09:56So we have two different scripts being written. We've got Ernest, our script writing agent, all in there. We keep track of this.
10:01We can add another agent in a new empty repo, and then we have all of our agents here assigned to different tasks related to YouTube. And the way it does that is if I just go command t, it's just doing the the Claude dangerously skip permissions command. This is just a terminal here, and I'm already logged into my Claude subscription.
10:17So it automatically just opens this up. So if you were to open this up yourself, you just need to auth in your account one time, uh, and it'll allow you to do the same. At the end of the video, I'll get into onboarding and actually setting this up from scratch to make sure that people have access to this and can start this up themselves.
10:31But for now, this is looking good. And this is the beauty of Fable, by the way. You give it feedback and it just does it.
10:36There have been no revisions off camera. All it took was, what, three iterations to get here. It is very fun working with such a capable model like this.
10:44Now the problem with this is we're only able to access Claude code like I just showed. And the whole point was running any model I want as well as my Claude subscription. So how do I get any model I want accessible within our daemon agent application?
10:56Well, honestly, I don't even know. So we are gonna go back to Fable and we're gonna tell Fable, alright, this is looking good. Everything is approved.
11:02Now what I wanna do is have an ability to add my ChatGPT subscription as well and log in to, I guess, would be technically Codex and use GPT five five within this, as well as any open source model. And so I don't know exactly how we can set it up with these different open source models. I'm assuming the easiest way to do this is through an open router API key, but I want an ability to when I open up the application, press that new tab button.
11:27Below the tabs, there should be an ability to select GBT, Kolod, of course, then OpenAI logo, then some popular open source models like Kimi k two seven, MiniMax m three, and GLM five two with their logos as well. So I can just click those and it will automatically open a session with that open source model.
11:45And so I think to get those open source models, the easiest way to do this is to just use an open router API key and we can just select all of these. I'll let you be the judge of all of this. Go ahead and spawn agents and do this next.
11:57Alright. Our new build is live. So we have an ability now to choose from different models, Claude, OpenAI, KAMI, MiniMax, GLM.
12:06And we should have little indicators, little icons to actually choose from these. So let's check this out. Okay.
12:12Looking pretty good. We've got the actual icons here to choose from. Minimax can't actually see that logo.
12:18Change that. GLM 5.2. Not sure if that's the right logo.
12:20Need to change that. But what we can do let's see if this actually works. Give me k.
12:25Okay. So nice. We have this enter your open router API key.
12:27So I'm gonna do that now. This should work the exact same way once you download this application, which again, we're gonna get into how to actually do that. But if I just go here, I go to OpenRouter right now, sign up for OpenRouter if you haven't already, but I'm going to assume that you're signed up.
12:40Go to workspaces, API keys. I just blurred that out, but I went create new API key, hit create, make sure you've got some credits added. And then I can go back to our application, press give me k, paste in my open router API key, save and launch, and then it automatically will do this to where it will connect.
12:57So we're using the Anthropic base URL. So this is loaded in the Claude code harness. This is the easiest way to do it.
13:03And I don't wanna get too in the weeds here, but what it's doing is it's just pointing to the OpenRouter API endpoint. We have our OpenRouter API key that I just added, and then it's it's calling this model moonshot k two seven via open router and then I can say, hey.
13:15And it looks like I'm running into an issue here for one missing authentication header. Not sure what's going on there. I just went back to Fable, asked Fable to figure this out and I'll fix this quickly.
13:23Also while Fable's at it, can you change the MiniMax logo to be white mode? Find the logo.
13:30Alright. Some quick revisions to the model pickers. We have updated logos here, and we should now have an ability apparently to open up any model here, Kimi k two seven.
13:38Hey, what model are you? Cool. I'm running with the Kimi k two seven coder model back end.
13:43So now we have an ability to use any model with any agent, whether it's Claude, Codex, or an open source one straight through open router, except now we've introduced another problem, which is if I close a tab, the agent's not gonna remember anything. We haven't set up the proper markdown files yet, instructing the agent on who it is and what it does, and have some sort of memory system and self improving system that makes this truly agentic.
14:06So that's what we're gonna set up next, and it's gonna be stored in this file section here. This is why I created this drawer here, to have a quick ability to click through and see exactly how we've configured this agent. So that's what we got next.
14:18We're gonna go back to Fable and explain, perfect. All approved. Now what I need is for us to build out agents files.
14:25So agents dot m d, a memory dot m d, and anything else that Hermes does, I need you to go through the Hermes GitHub repo here, determine how they have their agents self improving, and then I want us to just borg Hermes and bring it right over, and that will be our agent structure. So it's easier for me to just use something that already exists.
14:41I know Hermes has probably the best agent memory that they've configured, and the beauty of this is it's fully open source, so I can give this to my agent, have it parse through, find how they've structured the memory and kind of self improving system with their harness and use this for our own agents. Alright. So Fable grabbed the Hermes GitHub repo, brought over exactly how they structure this agent dot m d, user m d, memory dot m d, and this should now be set up in our application.
15:08So if we go back to our a d e, we should now have this set up. Yes. We have these files here, claw dot m d, and then we also have these agent files here.
15:16And so each one of these are a markdown file, which essentially makes up the agent. By the way, if any of this is confusing at all on the files that actually make up an agent or you just wanna learn more on how to build your own agents from scratch, you can check out this video. I'll link it above.
15:27It's a full course on building out AI agents. I'll also link another video above, is zero to your first AI agent in, like, ten minutes. It'll give you a complete walk through on all of this.
15:35But these are essentially the files that make up our agent, and we pulled this directly from Hermes. So we have our agent.m which just tells the agent who they are and what their role is. So you're Ernest.
15:43This is just a template, it looks like. So it's saying you're an autonomous coding agent, which is incorrect. And then here's the operating brief on how to actually access the rest of these files and update the rest of these files.
15:51For example, memory dot m d updating those as well, which it should do automatically after each session. But now we have an agent that is actually up and running. And what I would suggest you doing, because there is some optimizations, like I said, this is just a template.
16:05So what you can do is just open up your agent here and open this up in Claude and just say, hey. I want to update your agent.md. You are Ernest, a scriptwriter for my y t channel.
16:17Need you to help me with hooks and and outlines. And so it's updating its agent.md, and then you can continue to chat with it. And as you work with it, it's just gonna get better over time.
16:27Alright. So we now have an agent configured with the proper markdown files from an agent dot m d, identifying what it is to a memory, and this a d e is about built. So the last part of this is actually showing you how to set this up yourself because I've been building this with Fable, but we now have a public GitHub repo for you to access and download yourself.
16:47So that's exactly what I'm gonna walk through next. Okay. So I need to make this accessible in a public GitHub repo for people to download, make sure that you strip away all the secret keys and whatnot, organize this, make it nice and clean, have a walk through guide, and it should just download this Electron app.
17:01Right? Do we need to do anything with the Electron app to get this set up properly? Make sure this is super easy for people, and we go through that exact same onboarding that we had previously where we can, you know, create categories, create agents, add profile photos, all of that.
17:14Okay. So Fable has set up our GitHub repo and an ability to download this application yourself. So we're gonna do that now by going to this link here, which I, of course, will include in the description.
17:26We've got this GitHub repo, and I can install this. So download the signed d m g from the latest release, open it, and then just click this d m g here. It'll download.
17:36Also, sorry, Windows users. This is only for Mac, but feel free to still clone this repo and you should be able to with Claude, build out a proper Windows application relatively easily. Okay.
17:44And then just open this download here. Here's the DMG. My nice little icon here and just drag this into your applications and go to applications.
17:52Look for a d e here. Here it is. It's gonna ask you to open it, and you will see a blank screen exactly like we saw when we first set this up.
18:00So you can create a new team here. Test. Add a photo.
18:03Let's just call this YouTube again. Add a photo. Create category.
18:06Okay. Team. A bunch of these are just agents I was experimenting with.
18:10They're creating out, so you can ignore those. They should be blank for you. This, like, the whole thing should be blank for you.
18:14And then all you need to do is create an agent. So we're gonna say, Ernest, use a random image. There we go.
18:20Simmons bench badge. Create agent, and it'll just automatically open up Claude. Then if you want to open up, say, Codex or anything like that, you can just click there.
18:26You'll ask to sign into your product subscription or your OpenAI subscription, and then you'll be able to access that as well. So that is set up. You can see it really simple to do.
18:34We have agent files here and work with your agent on what you want the agent to do, but that's how simple setup is as well. Now I could keep going with this and building out more agents because we've really only scratched the surface on the different categories, the different types of agents that you could build, but that is beyond the scope of this video.
18:48Video. So what I'm gonna do is I'm gonna create a separate full walkthrough on building out more of these agents yourself and turning this into a proper quote agentic operating system. And in that video, I'll go through every agent I actually use from script writing to ones that help me with content, newsletters, all of that.
19:03I have a ton. That was also the purpose of this video was creating an ADE for myself to run all of these agents, to actually orchestrate all of these agents. But for now, this is more than enough to get you started building your own ADE and start getting your agents running completely for free using your existing subscriptions.
19:17Anyway, that is gonna do it for this one, and I'll see you in the next one.
The Hook
The bait, then the rug-pull.
AI agents are everywhere, but running more than one becomes its own kind of chaos, scattered chat windows, different repos, no shared system. This video is a real-time build log of the fix: a single Mac app, built almost entirely by describing what's wrong with each version.
Named ideas worth stealing.
02:23model
ADE layout pattern
- Categories (left rail)
- Agents (under category)
- Session tabs (top)
- File drawer (agent identity files, right)
The structural layout converged on after reviewing CMUX and conductor.build: group agents by category, open multiple tabbed sessions per agent, and surface the agent's markdown identity files in a drawer.
Steal forany internal tool that needs to manage multiple parallel AI agent sessions without a cluttered single list
14:57list
Hermes agent memory structure
- agents.md, role and operating brief
- user.md, user-specific context
- memory.md, self-updating session memory
A three-file markdown convention, ported directly from the open-source Hermes project, that gives an agent a persistent identity and memory across sessions instead of resetting on every new tab.
Steal forany custom agent harness that needs cross-session memory without building a bespoke system
How they asked for the click.
VERBAL ASK
17:14next-video
“I'm gonna create a separate full walkthrough on building out more of these agents yourself and turning this into a proper quote agentic operating system.”
Soft CTA, teases a follow-up video rather than pitching a product; also links two prior videos (an agents course, a 10-minute agent quickstart) mid-video.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
