02
The source line
The conceptual frame follows Addy Osmani’s Software Factories, Light and Dark and Dex Horthy’s work on why harness engineering is not enough. The autonomy ladder follows Dan Shapiro. Harness engineering follows Birgitta Böckeler. The field snapshot synthesizes primary and reported material from Stripe, Spotify, Cursor, StrongDM, METR, Thoughtworks, Sonar, and the major agent and specification-tool projects.
Volatile claims are dated and marked. Vendor figures are never presented as independent measurements. This is a July 2026 snapshot, not procurement guidance.
03
What came from our factory floor
The case study comes from building and running four local factories: first-factory, demo-factory, visual-novel-factory, and website-factory. The most useful evidence came from the website factory’s three-part verification model and its retro loop.
- Floor proves mechanical validity but never pretends to prove taste.
- Independent judges review visual and editorial quality in bounded rounds.
- Feel remains human because it names the residue those judges miss.
- Retrospectives measure what the human caught after the machines passed.
04
Image provenance
The home-page factory hall is an original generated image. Its prompt requires one continuous hall split by light, exactly one human reviewer, practical machines, the site palette, and no readable text or branding.