{
  "id": 5377727,
  "title": "It cost $33 to build a virtual Union Square. Here’s what the agents got wrong.",
  "url": "https://urgent.news/2026/09/03/it-cost-33-to-build-a-virtual-union-square-heres-what-the-agents-got",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-03T17:34:31.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/ai-agents-3d-playwright/"
  },
  "original_language": "en",
  "account": "PhiloLabs conducted an experiment to gauge the capabilities of AI coding agents in creating a virtual replica of San Francisco's Union Square. Using 453 building footprints, 75 custom façades, 129 named storefronts, 220 pedestrians, 109 vehicles, including Powell Street's cable cars, the agents constructed a working Three.js version of the iconic location. The entire process required 8 million tokens and cost approximately $33 in API calls. To identify visual issues that conventional tests might overlook, the agents were paired with Playwright, which checked the scene against 34 predetermined camera angles and captured 147 comparison sheets. These sheets highlighted discrepancies between the AI-generated scene and real-life Union Square, such as incorrect proportions or misplaced storefronts. The agents' work was then reviewed by specialist agents, who produced nine reports on the reconstruction. These reports served as a checklist for subsequent improvements. The challenge of the project lay in the limited data available—geographic data provided locations but not façade details, while photos captured only visible angles, leaving room for interpretation. Despite these limitations, PhiloLabs managed to create a detailed 3D model from open data sources, with a significant portion of the cost attributed to efficient task division among subagents and caching of context. However, this experiment presented limitations when applied to more complex applications, as the use of screenshots for error detection proved insufficient for identifying issues buried in code or those triggered by user interactions.",
  "summary": "PhiloLabs wanted to see how far a group of AI coding agents could get building something where working code wasn’t The post It cost $33 to build a virtual Union Square. Here’s what the agents got wrong. appeared first on The New Stack .",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}