Jay's Brief

Thousands of AI agents built 450 worlds that are drawn live as you fly through

A person sits on a terracotta couch looking at a laptop on a coffee table, its screen showing a glowing golden doorway that opens onto a hilly landscape. Beside the laptop, a tall, leaning stack of colorful binders teeters.
How much this touches your life Creativity

Medium: a free, sit-back look at what AI agents can build with light human steering, and where a person still matters.

What happened

Ethan Mollick, an entrepreneurship professor who writes the newsletter One Useful Thing, gave Anthropic's new AI model, Claude Fable 5.2, a one-paragraph idea: an endless flight through 3D worlds linked by portals. Twenty hours later, 204 worlds were live online. Now there are over 450, built by teams of AI agents. By Claude's own estimate, about 25,000 agent runs went into it. And none of it is pre-made. There are no 3D models, pictures or sound files behind the worlds. Your browser builds each one from code while you're still flying through the one before it, and makes up its sound as it plays. Mollick tested the model before its public launch and calls it "almost" AGI, or artificial general intelligence, which he defines as AI as good as most human experts at most computer tasks.

What's actually going on

Mollick didn't build this himself. His prompt asked for teams of AI agents (a reminder: AIs that work through a job step by step on their own). By the tenth day, a lead agent and 28 teams of agents were building worlds in parallel. Most of those 25,000 runs were one small job each, like judging a picture.

Why it matters that it's AI: one AI model can run many copies of itself at once, each on a small job. That lets work move far faster than one person could. Mollick supplied a list of places and a few feature requests. He says everything else, including the creative ideas, was Claude's, like hills sewn from denim with a mile-long zipper coming undone.

The weak spots showed up too. Mollick says some early worlds were uneven, and three separate teams each made a cyclone of circling animals. He told them to stop building and polish. The agents decided each new version had to win a blind vote by judge agents on twelve test screenshots, and teams started tuning worlds just for those screenshots. The lead agent banned that and wrote rules. After twelve days the rule book had 87,500 words, and output fell from hundreds of worlds a day to about a dozen. A separate "consultant" team of agents got it back to thirty or more.

Mollick's main point: the AI checked its own work for errors, but never asked whether it was too long, repetitive or no fun. He had to notice that.

The word you'll hear

AGI

Short for artificial general intelligence, AI as capable as most human experts at most computer tasks, but there's no agreed definition, so when someone says it's close, ask what they mean.

In your life

My take: this is a fun, free way to see what AI agents can make now, and where a person still matters. If you use AI at work, Mollick's lesson probably applies to you too. It can catch mistakes but may not notice when the work is dull or bloated. That part is still your job.

Be careful about

Left alone, the agents repeated themselves and buried the project in rules. Mollick stopped a separate business test before letting the AI contact real customers, worried it would take money and deliver poor work. He also says no human has fully checked a research paper it wrote.

Where to get it

The Portals site is free, runs in your browser and needs no account: portals-endless-worlds.netlify.app. It works best with a real graphics card but runs on a phone at lower quality. Try it with sound on (press S). The source doesn't say when Fable 5.2 reaches everyone or what it costs.

Go deeper

Fable 5.2 and Jagged AGI
One Useful Thing, November 9, 2026

Also checked against: GitHub