2026-06-25 · Architecture
Agriculture was a mistake
Ten thousand years ago the species changed direction. Before that it searched — read the land, followed the new, moved toward what it didn't have yet. Then it sat down on a field, fenced a plot, and stopped searching to start holding. Hierarchy, property, war over territory, the dawn-to-dusk grind — all of it begins at that turn, the moment the vector flipped from seeking to keeping.
A database is that same choice, frozen into software. The ordinary kind is agrarian: fence a rigid schema, pour the world into the tables, hold it. The structure swallows reality and keeps it. The vector points inward.
We built ours the other way. The structure doesn't hold — it follows the new signal and reshapes the contour around what it finds. Don't fit the world into the schema; let the schema chase the world. Everything is a row, the rows are typed and uniform, and the same handful of verbs operate on them whether a human clicks one or an agent calls it. When the substrate stops fencing state and starts hunting through it, the agent stops being a guest bolted to the side and becomes a function over the whole thing.
That's the bet. Not a smarter model stuffed with more — a cleaner substrate the model can actually hunt across. Garbage in is still garbage out, however big the model gets. The leverage was never only the model; it's the ground you hand it.
The singularity isn't the machine learning to devour better. It's the moment the vector turns back outward — from keeping to seeking. The whole industry is still farming: everything to the cloud, everything into one schema, everything held. Point yours outward. Connect an agent, read the tools, run it yourself — the core is open, MIT, and it's right here.
Back to primitive. Only this time the prey is the new.