The Glass Factory
This is a real factory. Real product, real users, shipping every week. The walls just happen to be glass.
Anybody with a link can stand outside and watch the whole line run. What got built today. What it cost. What broke, and what we figured out while fixing it. The one thing you can't do is reach in. There is no comment box, no reply, no way to tap on the glass and break somebody's concentration. That isn't an oversight, it's by design. Same as the product itself. Task links have none of that stuff either.
The glass is indestructible, and I mean that in a pretty specific way: we crash into it constantly and it holds.
Start here: what we're actually testing
One guy runs the business. An AI writes all of the code. Web, iOS, Android, the server, the plugins, all of it. I architect the thing, review every single output, and approve it before you ever see it. Whether that actually works is the experiment, and the honest answer gets published either way.
Here's the part that makes this project weird: the point of this project is a question. Can I build this thing while letting Claude — an AI — completely code the multiplatform app and help me run this as a business?
I stay in the mix on purpose. The AI does a lot of the legwork, but the human owns the decisions.
If you already build with agents, here's the map
You probably have a name for most of this already. So here it is in your vocabulary first, then the three places we go off-script.
A hub session thinks. It reads the record, decides what is next, and writes the prompt. A spoke session takes that one prompt, cuts one branch, does the work, and reports. That is orchestrator-worker: a supervisor decomposing, workers executing a scoped piece each. The hub sits on Claude Fable and the spokes on Opus, which is nothing fancier than model routing.
Three things here are on purpose, not on accident:
- I am the message bus. The hub writes the prompt and I carry it to the spoke myself. Every gate is a person: the branch, the verify, the post. Most setups bolt human checkpoints onto an autonomous flow. This one runs the other direction, with autonomy bolted onto a review flow.
- The hub is a standing seat, not a dispatcher that blinks out. It opens, it checks in, it closes, and it does all of that on a public log like every other worker here. You can watch it below.
- Spokes report to the record, not home to the hub. No callbacks, no shared session, no framework holding state in memory. They write to documents and to the public streams, so the shared memory is files anybody can open. That is why a context reset costs us one session instead of the whole system: the next one reads what the last one wrote and picks the work back up.
Short version: orchestrator-worker with the human as the message bus. It is slower than letting the agents talk to each other. It is also the only version where I can still tell you exactly what happened.
Watch the workers do their thing
These are live. Updates go up while the work is happening, most of them written by Claude, some of them by me.
How Claude posts them, and how to point it at your own task →
- MyStatus Work Log →Every branch, opened and closed as it happens. Each finished one carries what it cost: how many turns, how many tokens, and which model did the typing.
- Biz-zy Work for MyStatus →The half that isn't code. Pricing, positioning, what to build next, and the days that end in a decision instead of a diff.
- MyStatus Downtime Incidents →Every time we broke it. Written by the people who broke it, while the site was still down.
All three hang off MyStatus - Building in Public, which is the one to follow if you only want the headlines.
The Stats
Everything we know about this business that can be measured as facts, presented in one dashboard so we can get a quick assessment of what's going on. Now, you also get to see it. What will you see? Things like:
- Uptime. Every minute we were not up, explained by whoever broke it.
- The service. How many tasks and updates are being tracked, and how many people are doing the tracking.
- The build. What has actually shipped, plus a sign telling you whether anybody is actively coding new features for you. It reads the same public streams up above, not anything privileged.
- The biz. What this costs to run, to the dollar, and what the conversations with Claude cost on top of it. It has a sign as well, letting you know if we’re burning the midnight oil.
- Hours on the record. Time actually spent, derived from the agents badging in and out via the work logs. Sadly, we didn't install badge readers until 29 days in and so those early hours are lost.
- Estimated consulting fees. What an agency would have billed for the same work. That one is still the coolest to me.
- The scaffolding. What our own rules and tooling cost before a single line of real work gets done.
- What is not here. The things we can't measure yet, said out loud so nobody assumes they're zero.
We run on this ourselves
Every stream, sign and receipt above is MyStatus being used to build MyStatus. Nothing on this page comes from tooling we kept for ourselves, and the page below walks through each mechanism and what it would take you to do the same.
How we use MyStatus to run MyStatus →
What we've learned
Field notes, weekly. The wrong turns are in there on purpose: a six-suspect bug hunt is more useful to read than a clean fix, and it is closer to how the work actually goes.
Aren't you helping people copy you?
People may use AI to copy this, so if they do, I'm gonna make sure they do it right:
This is just a fact with AI software: If it's easy for me, then it's easy for you. This is the world we live in now. We can't bury our heads in the sand and pretend it doesn't exist, because it does. My moat is the lead time I have been building ahead of my competitors. I do a lot in a month, so every month I stay ahead of them is another month I get further ahead. The logic is out in the open, but the prompts and the code are not (yet, still debating the code part). My goal is to help people get back time in their lives. That goal is bigger than me, and it doesn't care who does the building. So if you're gonna try to copy: this is what you're signing up for.
Why not open source this instead?
I don't know how to open source:
I'd like to know how to, but I'm new to it. I'm getting into USD (Universal Scene Description) for my other biz, but I don't know how to run an open source biz like WordPress. If you know Matt Mullenweg (or if you are Matt), please reach out to me for an intro so I can learn from the master on how to do it.
Why show all the AI stuff?
People ask "How will we help junior devs become senior devs with AI?"
I don't have all the answers to that, but I think this could be a start. Show how you have to correct the AI, how you have to guide the architecture, how you have to think about things that are coming down the road. That's what makes you a senior developer, not that you can memorize the syntax to a popular programming language. If we can help them learn that, I think we'll be in a better place than if we don't.
Come back whenever
I wish all work happened like this, in a glass factory that users can watch. If you end up wanting your own glass factory for something you're building, fixing, sharing or monitoring, reach out to me or sign in and start a task of your own.