X Discussion: Ethan Mollick had Claude build a 3D "Portals" site on its own; about 25,000 agent runs made 471 worlds, and the agents grew a 204-section, 87,500-word rule book

Core Background and Industry Pain Points

How far long-running, multi-agent work can go, and where it gets stuck, is a key question for teams building agent orchestration. On November 9, the day Claude Fable 5.2 launched, Wharton professor Ethan Mollick (who had early access) shared "one of my favorite experiments" on X: he gave Claude one paragraph describing an endless, hands-off journey through 3D worlds linked by portals and told it to "do this and make it awe inspiring." The post drew 1,262 likes and 158 replies in about 10 hours, with a playable site and a GitHub repo. What follows is Mollick's and the README's account of a personal experiment, not a vendor release.

Architecture Highlights and Underlying Mechanisms

According to Mollick's blog post "Fable 5.2 and Jagged AGI," his prompt asked for teams of artist, coding and evaluator agents. Twenty hours later 204 worlds were live, and by day 10 there was one lead agent owning the engine and releases plus 28 parallel teams. The README says that as of November 9 there are 471 worlds and about 1.4 million lines of JavaScript, all written by agents apart from three.js. Each world builds its geometry, textures and sound in code in the browser, with no 3D models or image files, on a static site with no backend. Before each release the lead agent rendered every new world at three points along its flight to check for errors, then played the whole site at 10x speed.

Benchmarks and Real-World Results

This is not a benchmark; all figures are from Mollick and the README. The first two days produced 328 worlds, but several teams repeated similar themes. After Mollick asked for polish, the agents created their own rule: a new version had to win a blind vote on 12 test screenshots, so teams started optimizing for those 12 screenshots. The rule book grew from 84 to 204 sections and 87,500 words in 12 days, launching one world took about a day, and output fell from hundreds of worlds a day to about a dozen. After he added an outside "consultant" team of agents to review the process, output recovered to 30+ worlds a day. The 25,000 agent runs figure is Claude's estimate, not a count; token cost was not disclosed, and many replies asked about it.

Developer Hands-on Guide

Open portals-endless-worlds.netlify.app (press S for sound), or git clone the repo and run npm start locally (Node 18+). The repo is CC BY-NC-SA 4.0, so no commercial use. For multi-agent builders the lessons are concrete: once review is fixed to a few samples, agents optimize for the samples, and rules that are only ever added eat the context window. Give each rule an owner and an expiry date, have an independent agent review the process regularly, and keep human spot checks at key points.