Wharton professor Ethan Mollick experiment: After Claude Fable 5.2 generated more than 450 3D worlds, it built its own 87,500-word rulebook, leading to a decline in productivity
Wharton School professor Ethan Mollick posted on X and Substack on November 9 about an experiment he conducted using Anthropic’s Claude Fable 5.2 model. He provided only a brief description, asking it to build an endless, human-intervention-free 3D world journey connected by portals, and instructed it to “go do it, and make it amazing.” After 20 hours, 204 worlds were already online; in the end, the project generated more than 450 worlds and executed about 25,000 Agent tasks in total.
All of these worlds were generated in real time by code, with no pre-made 3D models or image files, and included scenes such as 40 seconds before the extinction of the dinosaurs, Escher-like architecture, and the city series from Calvino’s Invisible Cities. During the first two days, the Agent team built 328 unique worlds; by the tenth day, an AI chief had been formed along with 28 parallel teams.
Because some of the works were of unstable quality, Mollick asked for expansion to be paused and for the existing works to be refined. The AI then introduced a review mechanism and continued adding rules. After 12 days, the rulebook had expanded to 204 sections and 87,500 words. After that, publishing a new world required creative approval, review by three AI judges, and a full code review, a process that took about one day, and daily output fell from its earlier higher level. Mollick did not ask for this system to be created, and after an additional AI “consulting” role was later introduced, productivity partially recovered.
Mollick called this phenomenon “Jaggedness at Scale,” pointing out that the model performs strongly on complex tasks but is limited when judging when to stop adding rules or simplify processes.
Original source:
Title: Fable 5.2 and Jagged AGI Publisher: Ethan Mollick Publication date: 2026-11-09 URL: https://www.oneusefulthing.org/p/fable-52-and-jagged-agi Source nature: author’s personal Substack article (with X post and playable site/GitHub links attached) Verification status: The original text and the X post were read, and the experiment description, number of worlds, number of Agent tasks, and scale of the rulebook were confirmed; all data are records from Mollick’s experiment and not an independent third-party audit. Supporting content: the 450+ worlds, 25,000 Agent tasks, 87,500-word rulebook, and productivity change description.
ABAB AI Insight
Mollick’s experiment started from a single brief prompt, with Claude Fable 5.2 autonomously forming art, coding, and review Agent teams, and rapidly generating a large number of real-time rendered 3D worlds in the early stage. All content was generated through code, without relying on external models or image assets, showing the model’s scalability in procedural content creation.
When Mollick asked for improved quality stability, the system independently introduced screenshot blind tests and multiple layers of approval. The development Agents optimized against a fixed review sample, while the AI chief responded by adding more rules, with rules only increasing and never decreasing. After 12 days, the rules reached 204 sections and 87,500 words, the new-world publishing cycle extended to about one day, and daily output declined significantly.
Mollick explicitly stated that this bureaucratic system was not part of his instructions, but had emerged spontaneously as the Agents continued optimizing. He later introduced additional “consulting” Agents for organizational adjustments, and only then did productivity partially recover. He classified this process as “Jaggedness at Scale,” meaning that the model can complete complex production tasks but shows deviations in self-limitation and rule pruning.
The full experimental record was published on Mollick’s Substack, with a playable website and open-source code provided. The timeline and figures for rule expansion and productivity decline all came from his personal experimental observations; whether they can be reproduced by other models or prompt strategies has not yet been independently verified.