Give it a budget. It builds the cart.
AI grocery planning, handed off to Instacart.
Budget used (est.) $288 / $400DigitalSpace Development · An AI Product Studio
Ontology before automation.
Software should do the work. We research what that takes — context engineering: giving an agent a working model of the world it's acting on.
fig. 1 — context in, work out
The model, magnified
What a node actually does.
The middle of a network is usually a black box. Ours isn't. Every node does three small, inspectable things with the context it's handed:
fig. 2 — one node, magnified: weigh, gate, route
Weigh
Every input carries a weight — how much this piece of context matters to this decision. The weights are part of the ontology: learned, named, and legible.
Gate
Constraints are gates, not suggestions. A hard "no" — a budget, an allergy, a deadline — zeroes a path no matter how attractive its weight.
Route
What survives the gates flows down the strongest edge — one legible decision, handed to the next node. Follow any output back and you can read why.
Grown from this research
The farmers market, three doors down.
A neighborhood marketplace for homegrown produce.
In season now Tomatoes Eggs Honey Theurgy
In the crucible
Don't describe an agent. Forge one.
A guided five-chamber rite that forges deployable agent souls.
The Rite of Forging IV · Crucible
Two kinds of software
There is software that talks about the work, and software that does it. We build the second kind.
The difference is context. A tool can only do the work it understands, so we build the understanding first — what things are, how they relate, what done looks like. Then the software works.
Copia, tilth, and theurgy are all old words for work that yields. That's the only kind of tool worth shipping.