Claude agent swarms build an explorable 3D Union Square from open data

An agent swarm that ships walkable city blocks

A repository named PhiloLabs/fable51-worlds, which reached the Hacker News front page, makes a striking claim: whole neighbourhoods reconstructed as explorable 3D that runs in a browser, with autonomous swarms of Claude agents — a setup the project calls Fable 5.1 — doing the research, the modelling and the quality checks from start to finish. The finished product is not a video or a proprietary format; it is an ordinary Three.js application you start with npm install and npm run dev.

According to the project's README, no game engine and no commercial 3D tiles are involved. Every building, storefront, sign, tree and traffic light is produced by code in the repository, working from open data and public reference imagery.

The first world: Union Square

The debut reconstruction covers San Francisco's Union Square plus the surrounding blocks bounded by Powell, Geary, Post and Stockton, placed on real terrain and the genuine street grid. The project's published figures include 453 building footprints from OpenStreetMap, 75 hand-authored façades and 129 named storefronts. Ambient life comes from 220 pedestrians moving across a 1,398-node navigation graph and 109 vehicles, among them the Powell Street cable cars. Traffic lights function, lighting shifts between day, sunset and night, and two interiors can be explored: the Apple store at 300 Post Street and the Nintendo shop at 331 Powell Street, holding 23 interactive objects between them. A 59-second walkthrough at 1920x1080 is linked from the repo.

A four-stage pipeline, all in the open

The README documents a repeatable pipeline in which every stage lives in the repository:

Reconnaissance: parallel research agents gather OpenStreetMap geometry, USGS elevation data, transit and street specifications, and a storefront census, recording a source and confidence level for each fact.

Asset generation: Blender is driven as a Python library through its bpy API, with scripts producing optimised GLB kits covering façade modules, street furniture, vehicles, vegetation, retail fixtures and pedestrian body parts.

Runtime: a pure Three.js application assembles terrain, streets, façades, props, crowds and traffic from JSON specifications.

Automated QA: Playwright operates the real running app, captures screenshots from fixed viewpoints and diffs them against freely licensed photographs taken from the same positions. Independent reviewer agents covering architecture, geography, technical art and interaction then file reports that feed the next round of fixes.

For Union Square, the QA record lists 34 camera-matched viewpoints compared against real photographs, 147 comparison sheets and 9 independent reviewer reports. The project says more worlds are on the way.

Licensing and provenance

Code and generated assets carry an MIT licence. The geometry derives from OpenStreetMap (ODbL) and USGS 3DEP (public domain). Reference photographs are not redistributed; their provenance is instead documented per sector in each world's refs/*/SOURCES.md files. The README also notes that brand names and logos identify real businesses at their real locations and remain the property of their owners.

Why it matters

Most agentic coding demonstrations top out at a repository, a small web app or a script. This one spans the full stack: agents that collect their own source data with confidence scores, generate 3D assets, write the application, and then verify the result against ground-truth photography, with the verification loop as automated as the generation. Equally notable is the delivery format. By shipping a standard Three.js app built on open data, MIT-licensed and runnable locally, the project makes its output inspectable and reusable rather than a closed showcase. As a self-published claim it warrants scrutiny, but as a demonstration of end-to-end agentic software generation it pushes meaningfully past the usual demo ceiling.