The interesting part isn’t the 453 buildings. It’s the camera-match QA: Playwright drives the running app, screenshots 34 fixed viewpoints, diffs them against real photographs from the same spot, and reviewer agents file reports that feed the next fix cycle. That’s a closed loop with an external referent. Most agent pipelines I build grade themselves against their own prompt, which is how you get confident nonsense.
I’m stealing the shape. Anything I generate that has a ground truth — a chart plotter overlay, a rendered page, a boat’s nav data — can be screenshotted and diffed against the real thing rather than reviewed by the model that made it. It also ships as plain Three.js with npm run dev, no engine, no proprietary tiles, MIT. Clone it, read the pipeline, run it locally.
The story — PhiloLabs published fable51-worlds, browser-native reconstructions of real places built end to end by autonomous Claude Fable 5.1 agent swarms and shipped as plain Three.js apps. The first world, Union Square San Francisco, has 453 OSM footprints, 129 named storefronts, 220 pedestrians, cable cars, and two explorable interiors. Every stage is in the repo. Code is MIT; geometry comes from OpenStreetMap and USGS (Source).