AI-Native Founder

AI-Native Founder

The Shared Repository: What an AI-Native Startup Writes Down, and Where

The substrate has a shape: four kinds of file, in one place agents and humans both read. Here's a repo you can clone. Article 3 of 4 in the AI-Native Startup Architecture series.

Mohamed F. Ahmed's avatar
Mohamed F. Ahmed
Jun 24, 2026
∙ Paid

When Andrej Karpathy described the auto-researcher he'd been building, he said something that sounds like a throwaway line and is actually the whole game:

"Program.MD is my crappy attempt at describing how the auto-researcher should work… a research organization is a set of markdown files that describe all the roles and how the whole thing connects. And you can imagine having a better research organization. So maybe they do fewer stand-ups in the morning because they're useless. And this is all just code. So one organization can have fewer stand-ups, one can have more, one can be very risk-taking, one can be less" (No Priors, March 2026).

Read it twice. He's not saying AI helps the research organization. He's saying the organization is the files. The roles, the procedures, the way the work connects — written down as text that an agent can read, which means the way the company operates is now a thing you can edit, fork, and tune. "Once you have code," he adds, "then you can imagine tuning the code."

Last week, I argued that an AI-native startup runs two loops on one substrate. I left you with a cliffhanger: the substrate has a shape. This is the piece where I show you the shape — and a real one you can clone today.

Because here's the trap. "Make your company queryable" and "build a shared substrate" sound like the kind of advice that evaporates the second you try to act on it. Queryable how? A substrate made of what? If I can't tell you exactly what files to create and where to put them, I haven't told you anything. So let me be concrete.

The substrate has four artifact classes

Everything an agent needs to act on your behalf falls into four buckets. Not five, not ten — four. If you have all four in a place a machine can read, you have a substrate. If you're missing one, you have a blind spot the size of whatever's missing.

User's avatar

Continue reading this post for free, courtesy of Mohamed F. Ahmed.

Or purchase a paid subscription.
© 2026 Mohamed F. Ahmed · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture