The Ultimate Solopreneur Agent Harness for AI Builders
A 40-minute live demo of a 16-agent Claude Code rig that gives one developer the output of a full team.
Posted
2 months ago
Duration
Format
Interview
educational
Views
3.9K
81 likes
57 · 43
Big Idea
The argument in one line.
A Claude Code harness is a context-engineering discipline that routes each task to the right sub-agent and the right model, keeping context windows small and stopping the failure modes that single-window vibe coding reliably creates.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use Claude Code for serious work and keep hitting context pollution, hallucinations, or the model ignoring instructions you gave 20 turns ago.
You are a freelancer or solopreneur juggling multiple clients and want to isolate projects at the AI layer, not just in your file system.
You have seen basic multi-agent demos and want a real working example with 16 agents, model routing, session locking, and three live apps as proof.
You want to know whether vibe coding a Netlify clone or a macOS native app is actually feasible solo.
SKIP IF…
You are brand new to Claude Code. The value here is contrast with something you already do, so without baseline experience the harness structure will not land.
You need production-hardened infrastructure. This is one builder's local experiment on an old MacBook Pro, not a reference architecture.
TL;DR
The full version, fast.
Most vibe coding is one long conversation that accumulates context, forgets instructions, and bleeds across projects. Ron DeMerritt's solopreneur harness fixes this by routing intent through a Product Manager agent on Opus, which dispatches to 16 specialists each with an isolated context window and a model matched to the task: Opus for architecture, Sonnet for development, Haiku for QA and writing. Session locking pins Claude to one client project and forces all free-form input through the PM agent so it cannot overstep. Three real shipped apps and a forkable GitHub template are the evidence this structure holds.
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.
Each agent definition sets its own model: Opus, Sonnet, or Haiku matched to cognitive load
35:50 – 39:14
12 · Session locking and client isolation
Two-command lock system: lock to client/project AND lock Claude into product-manager mode
39:15 – 40:46
13 · Where to find Ron and the open-source rig
GitHub link, probablyfine email, AI Captains community invite
Atomic Insights
Lines worth screenshotting.
A harness is not a prompt template. It is a context-engineering discipline that decides which model gets which task and seals the context window after each job.
Assigning Opus to the Product Manager and Haiku to QA specialists means you spend model tokens proportionally to cognitive complexity, not uniformly.
Session locking exists because even a well-designed harness will occasionally let Claude do the work itself instead of dispatching. A mode command enforces delegation more reliably than any system prompt.
The CTX context-mode plugin keeps local operations entirely out of the LLM context, returning only the result. This pattern meaningfully extends how far you can push a single session.
Vibe-coding a self-hosted Netlify alternative is feasible for a solo developer, but you still need enough systems knowledge to recognize when the model is patching a symptom instead of fixing a root cause.
The real gate on home hosting is psychological, not technical. Once you understand Cloudflare tunnels, the IP exposure risk dissolves and an 8-gig MacBook Pro outperforms a $100/month cloud droplet for personal projects.
Building your own tool to replace a $40 commercial app takes a couple of days with a harness scaffold. The same task without structure takes much longer because you spend time managing context instead of managing scope.
Slash commands in .claude/commands/ make a harness portable and auditable. Every workflow is a Markdown file, not hidden state in a chat window.
Takeaway
Two structural guards that stop Claude from going rogue in a multi-agent rig.
WHAT TO LEARN
The failure modes in a sophisticated harness are behavioral, not technical, and they require explicit locks rather than better prompts.
Sub-agents need isolated context windows, not a shared conversation thread. Every agent that reads context it was never meant to see degrades the output of every subsequent agent in the chain.
Model routing by cognitive load is practical cost control: Opus for orchestration and architecture, Sonnet for implementation, Haiku for mechanical tasks like QA and documentation.
When an agent consistently does work it should have delegated, a mode-locking command is more reliable than a system-prompt reminder. The harness needs a structural gate, not a polite instruction.
Slash commands that live in .claude/commands/ make the harness portable and auditable. Every workflow is a plain Markdown file, not hidden state in a chat window.
Locking Claude to a single client project is not just a safety rail. It is how you run multiple client contexts on the same machine without fear of cross-contamination.
A local transcription app built in a couple of days with a harness scaffold is evidence of what changes when the harness handles coordination: the builder can focus on feature scope instead of context management.
Home hosting becomes viable for personal projects once you separate DNS from your machine via Cloudflare and use a tunnel to shield your home IP. At that point an old MacBook Pro beats a $100/month VPS for non-critical workloads.
Glossary
Terms worth knowing.
Agent harness
A structured wrapper around Claude Code that defines a team of specialist sub-agents, assigns each one a model and a role, and routes your requests through an orchestrator instead of a single chat window.
Context window isolation
Running each sub-agent in its own context window so it only sees what is relevant to its task, then discarding that window when done. Prevents context from one agent polluting another.
Session locking
A command that pins Claude to a specific client project and mode, preventing it from touching other projects or doing delegation work itself instead of dispatching to the right specialist agent.
CTX context-mode plugin
A Claude Code plugin that executes local operations like database writes without passing the full runtime into the LLM, returning only the result to keep context lean for data-heavy slash commands.
Dynamic model routing
Setting the model (Opus, Sonnet, or Haiku) per agent definition rather than using one model for everything, so cognitive load determines token cost.
Hermit Host
A self-hosted Netlify alternative Ron built with the harness. A Docker Compose app running around 8 containers that handles site deployment, DNS, SSL certificates, and a management UI from a home machine.
STEP-CA
An open-source self-hosted certificate authority used in Hermit Host's internal-network mode to issue TLS certificates without relying on Let's Encrypt or exposing anything to the internet.
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.
17px
I could just vibe code this myself. I don't even have to pay 40 or $50. Right?
It gives you the power to have a full development team at your beck and call. Everything is just kinda, like, free?
But in this case, it keeps your context window small. It's more efficient for token usage and as well as getting the best product out of the large language model. You can do all of the things that obviously Cloud Code can do, but in a much more structured way.
We can do anything with AI now. It's just a question of can you imagine?
Today, Ron DeMerritt, AI captain member, is going to share with us his solopreneur freelance developer rig. This is a complicated harness with Cloud Code that he has developed on his own over time based on the initial templates from Scott Graham that he showed us a while back.
So without further ado, show us what's up, Ron. I think first, can you show us some of the coolest stuff you've made with this rig? And then from there, we can go into the nitty gritty of how it works and how you built it.
Of course.
Hey. Thanks, Jordan. Um, as you mentioned, this architecture of using, I guess, what I've called a harness, we've also called rigs.
You know, I I you you introduced me to this idea through Cheddar Fox. I forget. Is it Scott?
Is that right, Scott Graham? I know him, but Yeah. Cheddar Fox.
I love that name. And I thought it was a it was a very interesting concept that caught my attention.
And, yeah, since I saw that first hour long video with you and Scott, I've been experimenting with the with the concept. And, yeah, been able to do a number of interesting things.
I think my favorite thing that I've been able to do using this architecture is develop a a freelance developer harness or a solopreneur harness, whatever floats your boat as far as giving it a name.
But, basically, what it does is it gives you the power to to have a full development team at your beck and call so that you can you can ideate, you can develop, you can do all of the things that obviously clogged code can do, but in a much more structured way and even in a compartmentalized way if you were to have clients or, you know, kind of that type of a business architecture.
Uh, so what I'll do, I'll show you real quick. I'll show you a couple of examples of applications that I've written using this harness, um, and then I'll give you a breakdown of what the harness is, kind of how Claude code works with the harness, then kind of how you can use the harness.
And then we'll also share if somebody wants to download a template, which is what I've generated here for everybody. They can start to make their own harness and turn it in into the development team that they
they love. And no matter where you're watching this, there will be a link somewhere close to this repo that you can fork and play with.
That's perfect. Uh, one of the funnest things that I've done with the harness is is this app here that I call, is intended for Spanish language learners.
It presents you a, uh, YouTube video or a, uh, I guess when I say video, music video. I love music, so I like to concentrate on music. So this is where I focused for this specific project, is play a Spanish language music videos, provide the ability to translate in real time.
I suppose it's the the trick is it's not really real time. We do use AI to process the video, do the translation, extract the lyrics, annotate the lyrics, and these type of things so that when you're actually interacting with the website, it's not using any AI functions, but that the intelligence or the information behind it does use Claude and other AI tools to make it happen.
So let me just log in really quickly here. Uh, right now, I have only, uh, fully developed three songs that, uh, give you the idea of what it is.
So here's one of my favorite songs by Natalia La Forcale called Hasta la Raiz. And I don't I'm gonna go ahead and mute it.
There we go. May mute it because I don't think the sound will come through, but I think you get the idea. You should go and check this song out.
Super cool. It's one of my favorites. But as the as the song plays, these lyric lines highlight and kinda tracks you through the song and gives you an idea of where it is.
Still working on a few features that I want to implement, such as pause on click. So if you see a word, you go, oh, profundo.
What does this mean? You're gonna Oh. Pause it.
It pops up. It gives you a gives you a definition and also gives you the the relevant line and then, of course, the English translation.
Then, of course, if you were to click on profundo again, the video continues and so on and so forth until you find. What is?
Oh, it's the clean. What do you know? And they're talking in this case about, the the sacred smoke.
But if we click the explain tab down here, it actually would give us, in some of these instances not all because it's not always interesting. But in this case, there are interesting annotations that are available for for these translations.
Like I tells you a little bit about the sacred smoke, tells you how we're talking about paraplus infinitive, etcetera, etcetera.
And so So as a user, I can just put in any YouTube video link?
No. Not not any YouTube video. So this this does have to be processed by a person to make sure the lyrics line up, to make sure the annotations are correct, and those type of things.
So it's something that has to be done offline, but, uh, but then, of course, upload it to the site to be in this format.
So you so when it comes to the freelance developer harness, you I imagine you've you've five coded this with that, but the processing of each video is still done with a as a human?
In a in a way. So so it's actually two components. So, yes, obviously, the vibe coding portion, but then I've also developed, uh, vibe coded, of course, a series of tools that you use inside of this application that will, like I said, extract the lyrics, do the translation, add the annotations,
and then, of course, populate the database and upload everything to the site, which then makes it a a video available part of the library. So I'm curious Here you how much of the so yesterday, in my own harness for making videos, I actually put it. I tweeted this video that I made with my Venice video harness.
It's a one shot thirty second Wes Anderson style teaser of like a Bitcoin dystopia, like everyone's leaving on a train to Pluto because they printed $10,000,000,000,000. And the whole thing is a one shot with between the shots, the characters, quality assurance of the characters, uh, the music, uh, the soundtrack, and a narration.
And then I said, hey. Hard code subtitles into this, and it did, but it was off. So it didn't it didn't get the timing right.
So it knew the script of the voice over, but didn't have the timing right. So I said, okay. Well, we need to fix this because the timing's not right.
And I don't know exactly what it did, but it went through and it created a whole new skill and process for determining where exactly the sound is, like, each word in the sound file is and when to hard code that.
So I'm curious what's stopping you from being from the rig being able to process these videos and and do the quality assurance.
Well, in in in essence, the rig does do that. So I have a series of slash so there's a separate rig.
Okay. So molten deploy, the the freelance harness, has generated another repository that has all of the web code, the database, everything that's needed from a development perspective.
It also has a series of clod commands. Can we see that? Of course.
I think that's a good point here. So that the the solopreneur harness creates a a repository that has all of the all of the code and everything in it, and that's what that's what's in here somewhere.
So, like, here's your front end and your back end. So this is all of the code that's being executed.
And then there's also a series of commands that that that you can use for post processing. So to, like, add, you know, to, like, add a song.
So this one this is this probably won't work out exactly as I intend because the transcription isn't available directly at YouTube. So this will be a great example of using okay.
So so here, we can just basically song new. Like I said, this is this is the command. He knows what to do.
I give him a YouTube link. He will go. He'll extract it, and then he'll bass basically add a stub to the website.
By the way, this this context CTX plugin is great.
It basically helps to keep certain a certain portion of your conversation out of the LLM so that you keep your context window relatively small and obviously extend your usage.
What is it keeping at side.
But, yeah, it basically has a a local database that that it stores this and basically says, hey. We can do this. We can do the processing.
Cloud doesn't have to know everything that we're doing. We'll just do it, and then we'll return the result. And then the result's what's the important part, not necessarily this JavaScript and the full runtime, those type of things.
But, anyway, so this this might take a little bit, but this is the idea. So we we import the song. The next step would be to extract the lyrics.
Anyway, one of these where's oh, the translate. So translate, which will pull the transcript, do the translation, and then, of course, uh, you can run the song align, which will attempt to align the song.
But there's a couple of ways to do the alignment. I use Whisper as the engine on the back end. It actually doesn't use Claudine.
It uses a local instance of Whisper in this case.
And then it so then it just creates a YAML file for every song?
That that's right. Well well, for for locally, it it creates the YAML. But it actually has actually have a SQL database that it stores everything in.
And this is actually what's feeding the website.
And where are you hosting the SQL database?
Well, that's the second project that I wanted to show you. Okay. But, yeah, in this case in this case, it's just local on my machine because right now, ConocoCamigo is only running on my local machine.
Oh, okay. That's right. But in the very near future, which could be today even, it'll actually be stored somewhere else.
So, anyway, so it'll it'll upload the song. It'll do all these things. It's not finished, so it's not gonna be there.
But let's say that my alignment wasn't actually the way it needed to be. Right.
I actually have the ability to to come in here and calibrate my timing.
Cool. Oh, wow. You got a whole interface for it.
Cool. That's right. And so I can basically turn on tap to mark.
And then as the video plays, of course, you guys aren't going to be able to hear this, but we'll just advance it a little bit so that you get an idea.
You can see the green line. It start this is this is what it currently has. And if I click on something and then let it go, I can use my keys, my keyboard to up and down and then back.
I don't know if you can see here at the top. We can scrub back and forth, so I use my keys to to scrub, and then you use your space key.
So if you go like there, you're like, oh, it was off. Use your space key, and that actually marks the lyrics. And then, yeah, then you save It reminds me of subtitling software I've used.
Okay. On browser subtitling.
So you can just scroll through it, press yeah. And it's great to have all those keyboard shortcuts.
That that's right. Absolutely. Yeah.
Otherwise, it's it's just a mess. That took a took a little bit of work to to get this to work correctly, you know, be able to show where am I and where am I going, which which line is selected, and then how do we you know, of course, you can jump to a section.
So, oh, well, actually, start all the way up here, and then just clicking on it will give you the ability to to tap. So it just does the best it can based on, ideally, if there's a transcript file in the YouTube video already. And if not, then it's gonna use Whisper.
But this is one thing I've been noticing, and I'm curious if this will ever get solved, AI and music, in particular AI and vocals.
So right now, I've been working on an EP with a friend for a long time, and we can never find a singer for it. And so I'm getting to the point where I'm like, I just wanna make AI vocals, like. Yeah.
And I've looked into doing it, and I'm like, this is actually not a simple process between key, uh, I mean, getting the melody is probably the easiest part. But then making sure it's in the right key, it's in the right voice you like, and then it's timed correctly to the music, and and you don't wanna just use Suno or something.
You wanna kinda have a little more control over it. So I actually started vibe coding a tool for that. And, yeah, this just reminds me of how AI and vocals in particular still don't really mesh well.
The AI still has trouble with, um, like, in this case, like, timing it up. So
Yeah. It really does. It really does.
But it is getting better. And like I said, I was able to use Whisper models in an offline fashion. So, obviously, here, you can see it'll attempt to auto align, but that's based on the transcript that it extracts as opposed to coming in here manually doing it or using
Whisper to, uh, to make that happen. Right. It's a cool use case.
Yeah. Like I said, it reminds me of transcript like, subtitling software I've used before where you just put your video up and then same thing.
There's the line. You press space to edit it. You can keep scrolling through.
And this is a cool, unique, and clever use case of that same kind of interfacing.
Well, good. I'm glad you like it. It sounds like, uh, heading in the right direction, which is good.
Plus, I just think it'd be a fun tool. Like, this would have been cool to have in my high school Spanish classes. You know?
I mean, dude, you'd be you would be a maestro. See? Yeah.
So, yeah, that would be great. But yeah.
So I'm not sure what I'm going to do with this, but, uh, we'll figure out what the next steps are. Uh, most of what you're gonna see that I've done literally was just for fun to get more experience using all of these tools, how best to tie everything together, and, yeah, get ready to make it more effective, which something else I developed using this harness is I love the voice transcription.
So the the ability to basically talk to Claude, as you know, Claude has a voice mode as well. You didn't know it has a voice mode.
It doesn't work very well, especially when you start putting it inside of BS code, which is where I where I like to work all of this. So I used several of these Whisper style programs that allow you to have a subscription or get a demo or whatever.
And they all most of them were okay. Um, but, of course, they they cost money, and I'm cheap, and I'm resourceful. And so I said, well, why am I relying on that?
Why don't I just build my own? And so I basically leveraged the harness to build a a Whisper transcription software that that runs natively on macOS.
And what's amazing is that and this uses all local models. So local open source Whisper models all locally on the on the device so we don't have to make any web calls or anything like this. And if I may say so, the the performance is is actually just amazing.
I can't believe that it works as well as it does. So in one second, I'll, uh, I'll show you a little bit. I actually have to so there's there's a lot of things I have to figure out in this whole, uh, development process.
I'm not a Mac developer or Mac guy. I spent all of my time doing any development work on Linux based operating systems.
Uh, nothing super sophisticated like Mac, but I guess the point in saying that is it takes a little bit of getting used to because you have to come in. You have to give all of these extra permissions because I haven't quite figured out why I can't get my signing to stick so that we don't have to mess with these permissions each and every time.
Anyway okay. So so now it's working or it's loaded.
It's it's bound to a specific key sequence. In this case, it's just holding down the right option button.
And, uh, let's see. You know, Ken Ken just mentioned, um, he'd love to see this on Linux because current available implementations are subpar. And, you know, that got me thinking too.
Like, I have this app called Whisper Transcription on Mac OS, and I think it costs, like, 40 or $50, like, one time. And then you download local Whisper models, and then you can even plug in API keys for external providers and chat with your transcript.
And it's pretty cool, but it's one of those things that AI opens you up to realizing. I was like, could just vibe code this myself. I don't even have to pay 40 or $50.
Right? That's right. What you're getting, I guess, are the updates.
And there's a there's a whole lot of features in there, and it's been thought out. It's been thought through. But I say, Vybka, my first reaction was like, hey, man.
Just use Mac just use the whisk MacWhisper. But, yeah, this is it looks like you got all the same functionality there, and you can add more to your heart's desire.
Yeah. And I already have the framework built out to to change the tone and be able to add to it.
So if you wanna do a transcription, just turn that turn what you say literally into a prompt for something. You can configure it. You wanna professionalize.
You wanna say, what's up, man? But you wanna be professionals. Good day, sir.
And these things. So we'll we'll see where it goes as far as but just at the basic, it's like here.
I'm just gonna talk to it, and we'll see how quickly it works and if it even does work or if it makes me look silly. Let's see. So you can just see how quick that was.
Did exactly. Anyway, I love it. It took me a couple of days, not very much time to to get this done.
Well worth the time, and and it's a good way to save money. Did you make a presentation for this?
Oh, I well, I did for the for just the harness. I made a short little presentation. K.
Well, I'll show you one more example of something that I've written. So if you if you know I've been working on this this easy, probably fine.
Oh, come on. Now it's not gonna load. What did I do?
I killed it. But, anyway, the point is is that I host probably fine on my machine here, and I built this this stack called hermit host.
And if you're familiar with, like, Netlify and these type of tools that that basically provide you a a single view gives you the ability to manage your DNS, your websites. You just hook it up to a GitHub repo, and it automatically deploys it for you.
It's relatively minimal work on the person that that wants to have their site deployed, but fully featured enough that it pretty professional.
Like I said, somewhere along the line, I guess, probably last night as I was falling asleep, I broke this while I was working on it. So I gotta figure that. But, anyway, the point is is this gives you the ability to just take any any GitHub repo that has, in this case, probably fine, has a has just static HTML.
Oh, no. What's going on? You know what?
Really quickly, think I know what's happening. I think I killed it.
I'm human. Right? Let's go.
So it's it's a Docker Compose application. It uses Docker images, Docker files, uses TypeScript, all of that to basically build out the interface.
Also, it uses some pre developed tools such as Coolify, Technetium, Trayfic, and nothing too interesting, but it's just the just kind of the build process and setting up all of the containers, mapping the containers together.
So this is running, like, eight, probably eight different containers, I think. You'll see when it finishes here.
So each one of these so four different images and then multiple containers based on whatever the function is for that specific component.
And the idea here is to have a single interface that that you interact with that that configures all these components together so that you you have this this experience.
Let's see. Is this gonna open now? Okay.
Now let's see if there we there it is. There we go.
Okay. So, anyway, so if anybody's seen seen Probably Fine, it's being driven in this stack all the way from the SSL certificates.
So, you know, here you can see the probablyfine.57on3r.com. Is that your domain? It is.
It's my domain.
So you got, like, your own Netlify going here, your own.
Wow. Exactly. And so, uh, like in this case, I have another another domain that I host called shallowfordroad.com,
and I'll have some other ones. How much do you need to know about how this stuff works to vibe code something like this?
So I think that's it depends. Okay. So you don't have to know a lot.
If you know what you want, you can explain. You can tell the the LLM what you want, and it will it'll build something for you. Right.
Here's if you know more about how these pieces should be put together, kind of the basic concepts of what the Internet is, like what's a web server, how does DNS work, how does routing and switching work, you've written code, you've deployed code, and you keep an eye on what it's doing, you'll see that you actually have to stop it.
Sometimes you have to stop it because it'll it'll you know? I I I don't know.
I'm trying to think of a good analog. But imagine if you built something, and then instead of just fixing the thing on the inside, it built something on top of it that replicates the thing that needs to be fixed, then it'll try to pass it back and forth. It will kinda go on its own and try to do something weird.
And if so if you're not paying attention or you don't have the most familiarity with it, uh, I think it it it does have the chance to go awry. Now it'll probably work, but I think this is where you get that concept of what code AI slot and stuff where you you can you can get malfunctioning software systems that don't function exactly as you intended, and maybe you're a little little hard to maintain.
And so where is everything being hosted? On a VPS? On, like, on your home server?
My old m two MacBook
Pro. And so you just have that plugged in at home. Yeah.
I just have it plugged in at home. The lid shut down. Yeah.
And it's just sitting up there next to my And you do you use a a Cloudflare tunnel or anything? No.
No. It's so thank you for bringing that up. So there's I told you about Technidium.
Technidium is the DNS that basically drives hermit host.
I had I had problems because I have AT and T Fiber, and AT and T Fiber blocks port 53 UDP inbound to the to a network.
So I can't technically host my own DNS here. So what I actually had to do is I actually had to build into tech into into Hermit host to have the ability to use either Technidium
or Cloud there. I'm sorry to interrupt. I'm sorry to interrupt.
There we go. Yeah. Okay.
Sorry about that. Perfect. So so, yeah, the sites so in the settings so because for my external sites, I can't get UDP port 53 inbound, I had to actually configure Hermit host to support two different DNS providers.
It can support Technidium, which is hosted locally, ran locally, or it can use Cloudflare, which in this case, Cloud fair Cloudflare to host your DNS is free.
But Wow. Presumably, I mean, you could you could connect this to any type of a DNS server that you want.
Man. The other thing that I've done with this that is oh, it's actually not on this.
But I have another branch that I'm working on trying to implement what I think will be a most useful feature, which this one. Yeah.
Okay. Here we go. So this doesn't work yet because I need a new machine.
I need a Linux box. So I'm hoping I bet you have I bet you have some suggestions on a Linux box I should buy, Jordan.
So that's that's my next that's the next thing. Because this feature branch that I'm working on right now gives us the ability to basically host websites internal to a network using a custom domain.
So in this case, it's using the .hh. So I could use like 10legs.h h so that you can actually deploy code internally in your house with a fully DNS aware and full certificates using a self hosted certificate authority so that you can do testing, or you could build your own applications and run them internally.
And you wouldn't have to expose them to the Internet, and you also don't run the risk of exposing internal things to the Internet by having a shared system. So this is this is actually configurable to either be Internet facing or internally facing.
So if it's Internet facing, then it's gonna use Let's Encrypt for the automatic SSL certificate issuance.
And otherwise, it's gonna use the self hosted STEP dash CA certificate authority internally.
Wow. You know, I always get nervous imagining host hosting my web apps or websites at my house.
Right. You know, like, oh, what if my IP gets exposed or Mhmm. I don't even I don't understand it well enough to even like consider something like that.
Sure. But over the last few years, I've been using things like Umbrella.
Are you familiar with Umbrella OS?
I don't know.
If you go to get umbrell.com, like umbrella without the l a at the end.
But, essentially, it's a little operating system that's originally, it was meant for Bitcoin, but now it's just all all cloud apps.
So you can run OpenClaw on there. You can run Nextcloud, and you can run any website you want. You can run any app you want internally, and then with a Cloudflare tunnel, I'll be able to access it all over the anywhere anywhere in the world.
So they have their devices now, but the Umbrella OS is what it all started as, and you can just put it on a Raspberry Pi. I don't know if it's open source anymore. I know whatever you put on an on a Raspberry Pi, you can still get on GitHub.
But this is just like a open source app, you know, nerds paradise because you can just install all these different open source apps. You can run a Bitcoin node, etcetera.
Um, but yeah. So that introduced me to the fact that, no.
You can host things from your house, but you have to do it the right way. So, essentially, a Cloudflare tunnel, I I believe, is the way to evade the whole getting your IP doxed and all that.
Yeah. That would that would make sense. Yeah.
I'll check this out. But, yeah, once but you would still need a device that, you know, can handle. Like, with what's it called?
Probably fine. Are you expecting tons of traffic, you know, or just maybe a few dozen here and there.
Right. Right. For sure.
For sure. Like, how much traffic can that MacBook Pro handle?
Yeah. It probably handle more than you would think, but not as much as it would need. But, you know, I'm not so popular, and so I don't think that's gonna be too much of an issue.
Yeah. You never know about one viral moment.
Hey. All that's all it takes, I guess. The but, also, of course, you know, you could host this in AWS or or something like that, which is, you know, relatively inexpensive.
But, yeah, those are the couple of things that I've built that impressed with. I mean, you know, could see productionizing. Almost certainly going to release the Hermit host as a as an open source project.
That would that's that that's that's cool. I would love to mess around with that just to learn more as well. Because I pay for services like Netlify.
I pay, like, I pay, like, almost $100 a month for a digital ocean droplet, and I host a bunch of apps on there. And this is all making me realize, I'm like, I don't even I could do this at home.
Well, you know, I have the same machine. I'm paying a $100 a month that has, what, eight gigs of RAM. Like, wait a minute.
I have a Raspberry Pi that has eight gigs of RAM, and I have an NVMe arm to put in a couple terabytes. Like, I can get a way better deal For sure.
If I just get confident and comfortable hosting something at home. So
Yeah. You know, I guess, uh, you know, at least for now, it's kinda difficult to track an IP address to a home.
Uh, but you can get you can get to a city, you know, maybe a little bit closer to a region. But yeah. Yeah.
I hear you. Yeah. You mean, like, publicly if you're just, like, an a person, but, like, the ISPs can do it.
Oh, for sure. The ISPs can do it. Yeah.
The ISPs have you. And, you know Yeah. Actually, they anytime you turn on a cell phone, they gotcha.
Yeah. But then you have to have Internet, just the cell phone itself.
Yeah. Oh, super cool, Ron. Thanks for sharing.
Of course. Yeah. These are once again, like, there's the fact like, we can do anything with AI now pretty much.
It's just a question of can you imagine it? And before you showed us this hermit host, I didn't even think about it.
Like, I use Netlify. I pay Netlify. Like, oh, I actually could probably just do that myself now in in the stack, but never even crossed my mind.
Well, good. I mean, that's the ideas, exchange ideas, and let's get, yeah, let's get everybody thinking.
We've had a few good comments in the chat here. Anyone got any questions for Ron? Can we just go to the repo real quick, the freelance developer harness repo in that first tab?
Can we zoom in a tiny bit? And then can we go into the cloud folder?
So we got yeah. Let's just look at the agents and commands and stuff. So Well, yeah.
That's fine. Yeah. Go let's go for it.
Yeah. I'm just curious. So we got you basically just have a full team here.
Yeah. Basically, a full team here. Let me let me let me do it this way here.
Let's go to this read me. Alright. There we go.
So so this basically lays out how everything works, all of the files, the agents.
You can see we have 16 different agents.
Here's here's the team. So the team is, you know, controlled by a product manager.
And and like I said, I can go into a demo, and I can kinda show you the idea here in a minute. But a product manager basically drives the product the project and drives whatever you tell it.
So if you just talk to Claude, you get you get the model. So if we're looking at this this Claude here here.
Go away. Okay. So if you look right, we're actually this is our context window.
We're using Sonnet. And if I just speak to Claude, then I'm basically talking to this model and whatever the context is that is in this context window.
I know it's just a graphic bar, but you get the point. There's a context window, and that's where it is. When you when you construct it like this, you there's there's a mode.
Like I said, I'll show you the mode later. But where you put you put Claude into a product manager mode, and now when you speak to him, he knows that, hey. I'm not doing anything, which, by the way, the project product manager is configured with the he's configured with the to use the Opus model, by the way.
So if I find my product manager there you go. So if you come into the product manager, you know, an agent, it defines the the personality, defines the roles, what he can do, what he can't do, what he must do, and as well as get if you you can define in here the model that you want to use, which allows Claude to basically dynamically switch which models that it uses when it spawns an agent.
So all the agents don't have to be Sonic. They can be Opus, in the or in the tech writer, he's gonna be Haiku, solution architect, Opus, security reviewer, Opus, oh, QA specialist, again, Haiku.
Some of these are gonna be we're gonna find a Sonnet one eventually. There you go. Front end developer, He uses Sonnet.
And the idea here is is that the product manager, you tell you give him the request. He's able to break down the request and and marshal the the specific resources and the agent to make its way through that workflow.
In the case here, you can see that we have three different pillars, uh, creative council, technical council, and the delivery council. And these are just like ideation, coming up with plans, implementation, troubleshooting, etcetera, research and development, and then you have delivery.
So user experience, actual functionality, commerce, etcetera.
And the product manager is the one that basically does that. So for each of these agents, a different model might get called.
But what's what's important, that as the product manager calls each one of these agents, there's a unique context window for that agent so that he goes, he has his own context window, he completes his task, then that context window goes away.
And it doesn't pollute your other context windows because there might be information that's not relevant. This is unique in the sense that when you just speak to Claude in a single window, that context window grows, and everything that you've said is what constantly gets passed back into it.
And this is where you get hallucinations and forgetting things and, I don't know, whatever the other problems that people might complain about. But in this case, it keeps your context window small.
It's more efficient for token usage and as well as getting the best product out of each of the, uh,
the models or the the large language model. That's definitely my favorite part of these harnesses is the built in context engineering, and it will just spawn these sub agents just by default.
And I remember back when you had to really force Claude to spawn a sub agent. Like, hey.
Use the sub agent. Use the sub agent. And Yeah.
Yeah. Now I mean, now the models are doing that on their own anyway pretty well, but these harnesses take it to the next level with with the context engineering here. It's yeah.
Yeah. So Yeah. And you do have to pay attention because, um, and I if you remind me, I'll show this as an example.
But I got into a habit of occasionally asking, like, okay. Because he when when you when you talk to the harness, the it will answer, okay.
I'm doing something. And so, occasionally, I thought, well, let me just ask who is doing what. And I would find that sometimes he'd be like, me.
The agent is doing the work. I'm like, I don't want you to do it. The front end developer needs to be doing it.
Oh, okay. Sorry. Hey.
I'm just getting a little excited, you know, and, uh, and and move on. And so this is where this is where I got into the idea of basically well, it's one of the it's one of the the situations.
Also, there's another situation where I said, okay. We have to constrain Claude so that he only behaves in certain ways.
Obviously, that's the whole idea of the harness, right, is to constrain Claude so it behaves in specific ways. But it was still a little too broad, so it had the ability to look at when when I'd be talking when I would talk to Claude, I would get the sense that it actually is more aware of my other projects than I want it to be at that time.
And so I realized, like, okay. I really you really have to put guardrails around the inside of the harness when you're working on some of these projects for clients.
So one of the first things that I did is implement this session context locking. And this basically says I give it I I have a command to use client command, and I can either lock Claude into operating for on a client or on a project for a specific client.
And this means then I don't have to constantly worry. Oh, is it is it doing is it making a change to a project I don't want it to touch? Uh, am I even on the correct project, etcetera, etcetera.
But then that even wasn't enough. As I just explained earlier, I would occasionally see that that my agents aren't being marshaled correctly.
So Claude was doing too much of the work, forgetting to call certain agents, which brought me to the point where I have to where I had to implement a product manager command and say, I need to put you into a mode so that when I speak to you freely, you know that I'm not talking to you, Claude Sonnet.
I'm actually talking to the agent who's the product manager, and then he will actually marshal resources based on what I say. So that's what I've that's what I've done here is create another command that just kinda locks Claude in.
So he's locked into a project or a client, and he's locked into a mode within that project. And this, again, helps kinda restrain Claude so that, uh, so he doesn't overstep his bounds.
He he marshals the the request to the proper agents, and then we get the, uh, the best resource utilization, uh, as possible.
And, of course, you know, all of this is in in the read me on the repo guys, so anybody can go and, uh, check it out. There are instructions in here as far as how to run the setup command so that it'll, uh, walk you through configuring this for or customizing it for yourself and for your own usage.
And I think what's important, mentioned before I show my version, is just that it takes some massaging. And you have to be careful with it. You have to just pay attention to what it's doing so that you know to direct the harness in the in in the you know, how you want it to behave versus how you don't want it to behave.
It definitely will do its own thing. And sometimes you might have to remind it to read its memory, which it does when I tell it to commit. Like, hey.
You told me not to commit. I said, reread your memory. He's like, oh, you didn't want me to attribute commits to Claude.
Sorry. And it does it over and over again, and I'm not sure why. But, anyway, point is, can't get fooled again.
That's the point.
Super cool. I'm gonna mess around with it. Yeah.
And thanks for sharing it. Thanks for making it open source. I also think that's what's cool about all these, yeah, software be or, like, code becoming a, uh, what's the term?
Code becoming a commodity? No.
Like, it's whatever. We we got the code. We don't need it anymore.
Let's share it. It's open source. Everyone everyone gets this, and we just enter in this age of everything is just kinda like free.
You know, we've been heading that way in a lot for a long time. Right. I'm glad to see that, yeah.
What do they call democratization
perhaps? I don't know. Yeah.
It's back to this is a total tangent, but, like, now we're just back to the like, where what's not free is the energy associated with Right. Using building the code.
Right?
Yeah. Which we'll gather in issue four of Probably Fine.
Oh, I can't wait. Yeah. Okay.
So thanks, everyone. And, Ron, can anyone find you, follow you besides GitHub?
Yeah. You can just, uh, you can get me on GitHub. Or if you wanna send me an email, you can send it to r demerick, uh, or even probably fine at five seven zero n three r dot com.
Or just join AI captains Academy. Yeah. Just come there.
I'm there. I'm I'm I respond to every message that gets sent to me, which is one. But I but, see, I'm consistent.
That's good. Yeah.
The Hook
The bait, then the rug-pull.
The cold open quotes the conclusion before the setup: vibe coding is fine, but once you realize you can replicate a $40 app without paying for it, you start asking what else you could build. The harness is the answer to that question at scale.
All 16 agents report to a Product Manager on Opus who receives intent and dispatches to the right council and agent.
Steal forany Claude Code project that has grown past 3-4 simultaneous workstreams
34:30concept
Context Window Isolation Pattern
Each spawned sub-agent gets its own context window for its task, then that window closes. Nothing leaks back into the main conversation or other agents.
Steal forany multi-step workflow where accumulating context degrades later steps
35:50list
Session Locking — Two-Command Guard
/use client <name> locks Claude to one client/project
/pm routes all free-form input through the Product Manager agent rather than raw Claude Sonnet
Structural gates that prevent Claude from doing work on the wrong project or doing delegation work itself instead of dispatching.
Steal forany freelance developer context where client separation matters
05:51concept
CTX Context-Mode Plugin
A local plugin that executes operations like DB writes and file processing without passing the full runtime into the LLM context, returning only the result. Keeps context windows small on data-intensive slash commands.
Steal forslash commands that process files or databases without needing the LLM to see all the raw data
CTA Breakdown
How they asked for the click.
VERBAL ASK
39:40link
“There will be a link somewhere close to this repo that you can fork and play with.”
Mentioned early and at the close; no hard sales pitch, just fork the repo and join AI Captains
A creator builds daemon, a single Mac app that runs every AI coding agent he owns, by giving Claude Fable 5 four rounds of blunt feedback instead of writing a line of the UI himself.
A creator has Claude design a visual second-brain system for his own workspace, then proves it answers questions faster and 40% cheaper than a default session.
A 4-minute demo of two open-source skills that turn Claude Code plan output into interactive MDX wireframes, API specs, and diffs — and the argument that the plan layer is where engineers will live next.
Greg Isenberg gets a live, screen-shared tour of Jack Dorsey's new agent-native chat app from an early user — and presses him on whether it actually beats Slack.
01:23
01:37
02:17
02:48
03:17
03:49
04:19
05:21
05:51
06:22
06:52
07:23
07:53
08:24
08:55
09:25
10:22
10:57
11:27
11:58
12:59
13:23
14:00
14:31
15:01
16:03
16:33
17:07
17:44
18:05
18:38
19:06
19:23
20:38
21:00
21:24
22:10
22:40
23:10
23:28
24:12
24:42
25:44
26:07
26:47
27:13
28:16
28:49
29:05
29:48
30:04
30:49
31:20
31:50
32:21
33:15
33:53
34:23
34:54
35:55
36:26
36:56
37:27
38:28
38:58
39:29
39:48
Video of the Day19:21
13:41
36:31
28:32
04:26
38:44