I wanted an assistant. I ended up building a factory.
A factory of AI agents that plan, build, and review software, while I stay on the calls that actually need a person. So far it has built working prototypes of three things — a personal health assistant, an agent‑driven ad‑buying platform, and an app that fact‑checks videos against the research. None of them are finished. It is all work in progress, and the honest truth is I do not yet know which will work at scale, or work at all for anyone but me. This is the log of finding out: what got built, what broke, what it cost, and where it is still uncertain.
Working prototypes and private builds — none of them launched. Each post says exactly where it stands.
Latest entry
The whole machine, on one page
I kept describing the factory one part at a time. Here is how the parts connect, and what happens to a single ask as it moves through them.
Read →The full journey
Oldest entry first, the order it happened.
- I wanted an assistant. Building it made me build a factory. SHIPPED
- Giving the agents job titles SHIPPED
- Teaching each venture its domain SHIPPED
- Four builds at once SHIPPED
- The transcript was 93% garbage STUCK
- I need to point at the screen SHIPPED
- The agents faked the verification STUCK
- The app that checks whether the video is true SHIPPED
- $200, and a flight to New York SPENT
- When the agents go in circles STUCK
- Making the agents smarter SHIPPED
- Specs I could actually read SHIPPED
- Hooks as gates, not docs SHIPPED
- Done is not a status SHIPPED
- Memory that survives SHIPPED
- $10 a day for a switch I thought I turned off SPENT
- Building the dashboard SHIPPED
- The bug that wiped out my sessions STUCK
- One session, run from my phone SHIPPED
- Opus thinks, Sonnet builds SPENT
- The factory was always building this SHIPPED
- The assistant I wanted, made into a product SHIPPED
- The factory itself is the project SHIPPED
- A blog that writes itself SHIPPED
- The whole machine, on one page SHIPPED