PhiloLabs rebuilt San Francisco's Union Square in the browser for $33

The New Stack reports that PhiloLabs used Claude Fable 5.1 agents to build a 3D reconstruction of San Francisco's Union Square, built as a plain Three.js browser app from real-world geographic data and reference images. Two hours later, the agents had produced a working scene with 453 building footprints, 75 custom façades, and 129 named storefronts, along with 220 pedestrians and 109 vehicles moving through it, including Powell Street's cable cars. The reported API cost for the run was about $33.

How the verification loop works

The interesting part of the experiment is what the agents did after the code ran. PhiloLabs divided the reconstruction among subagents that handled geographic research, building geometry, textures, storefronts, and other parts of the scene. A headless browser then moved through the finished app, stopping at 34 fixed camera positions, and captured screenshots that could be checked against photographs of the real Union Square taken from the same spots.

In all, the agents produced 147 comparison sheets, which made it easier to spot things that were technically correct but still looked wrong. A building might be in the right place with the wrong proportions. A storefront might end up on the wrong side of the street. Using the same camera positions on every pass also made it easier to see what had changed from one run to the next.

The reviewer agents

Specialist reviewer agents then examined the material, some focused on architecture and geography and others on technical art and interactions. Together, they produced nine reports on the Union Square build. Those reports became a list of fixes for the next pass, which the development agents worked through before rerunning the scene.

The company's own README states this as a hard rule: builders never grade their own work, and every world is published with its own QA report, defects included. PhiloLabs also acknowledges the limit: another agent checking the work does not guarantee the decisions are right, especially where the source material is incomplete, because the reviewer can miss the same thing the first agent did.

Why the token budget stayed small

Eight million tokens is a lot of activity for a single application. The $33 figure came from splitting the job among subagents, running tasks in parallel, and reusing cached context rather than having one agent work through the whole project from scratch every time. The output is a repository, PhiloLabs/fable51-worlds, under an MIT licence, with a second walkable world already published for Higashiyama in Kyoto.

The loop works beyond 3D

The pattern applies beyond 3D. A verification loop that runs the finished thing and compares its output to a trusted reference is the kind of test that catches errors a unit test cannot see: a chart labelled correctly but drawn against the wrong axis, a dashboard whose numbers sum but were taken from the wrong table, a landing page whose text obeys the brand rules and still reads as generic. A team building any output an agent produces can copy the steps. Fix a set of viewpoints or inputs. Capture the output on every run. Compare it with a trusted reference. Have a separate agent write the list of fixes, so the same reasoning that produced the code is not also grading it.

Source

Primary source: PhiloLabs/fable51-worlds on GitHub. Reporting: The New Stack.