How to Build an Ontology for AI Agents: Your First Manifest

An ontology sounds like an enterprise project: consultants, workshops, a diagram nobody maintains. At founder scale it is something else — database schemas plus discipline. It is not designed on a whiteboard; it is harvested from what your business already uses. Five moves, one day.

1.0The five moves

Harvest · Home · Edges · Gate · Version

01Harvest

Harvest, don't design.

List the object types your business already uses: customers, offers, projects, reports, standards. Don't build a new world — make the existing one explicit.

02Home

One home and one name per type.

Every object type gets exactly one permitted place and one naming convention. This kills the single most common source of drift.

03Edges

Write down the connections that matter.

Which object feeds which process? An offer hangs on a customer, a report on a decision. Only record the edges that answer a question.

04Gate

Define one gate.

The point where a human approves. The model produces and flags its own uncertainty — it never approves itself.

05Version

Version from day one.

Every change dated and explained. Without a changelog, your system cannot be reconstructed six months from now.

Diagram of the five moves: harvest, home, edges, gate, version — connected left to right

FIG 1.0The five moves as one diagram — harvest, home, edges, gate, version.

2.0Worked example

One run through the ontology: the CI PDF

A report becomes a delivered PDF in corporate identity — as a gated action along the ontology. The PDF was produced by exactly this pipeline. The artifact proves the process.

Ainary · Workspace › one run, five objectsSanitized demo

ObjectReport

lives in exactly one place, named by convention

StandardCI

versioned; governs typography, colors, layout

ActionPDF export

runs against the standard, never freehand

GateApproval

human · before every write

AssetCI PDF

filed, linked, traceable to the run

FIG 2.0— The same chain carries sales, finance and content. Only the objects change.

Where does ontology-first not pay off? Honest answer: for one-off tasks you will never repeat and never need to audit, the setup does not amortize. The rule I operate by: ontology first for everything you repeat and everything you must be able to trust.

See the ontology run, live

The sales pipeline from my own operation runs on exactly this manifest — gated, logged, end to end.

See the loop run

The argument · Structure over Scale 01 · Substack

Alex Karp Thinks in Nations. I Think in Databases. The Logic Is the Same.

Why structure and governance beat raw scale — the essay behind this walkthrough.

This article is the applied companion to the opener of the Structure over Scale series. Produced with an AI-assisted content pipeline and a human review gate (C5). The ontology manifest described is real and operational (v0.4.5, July 2026).

Get the next one before it publishes

Essays and research land on Finite Matters first. No noise, no schedule — only when there is something worth reading.

Subscribe on Substack →About Ainary