Ask an AI to check its own work and it will usually tell you it's done. We asked Claude Fable 5.2 to build a 3D world entirely in code, twice, from the same prompt. Once it checked its own work and called it finished. Once it had to get past a brand new Critic Agent every round, one that saw only screenshots. Then a blind Judge Agent compared the two.
In November 2026, Claude Fable 5.2 built the same late-night trading desk twice in Claude Code: once checking its own work (11 checks, about an hour) and once with a blind Critic Agent every round (26 rounds, almost five hours). A Judge Agent then reviewed both finished worlds nine times each, without knowing which was which. Up close, it found a major problem in the heatsink canyon in 100% of reviews of the self-checked build against 11% for the Critic Agent build, in the motherboard 78% against 0%, and in the graphics card fans 67% against 0%. The wide room shots went the other way.
Explore AI Agent Lab: Critic: zoom into matching shots of both builds, watch both tours, fly the world yourself, and download the free Critic Agent.
What did we build?
One prompt asked for a 3D world that runs in a web browser: a late-night trading desk with glowing monitors, a keyboard, and a glass-sided computer tower, plus a camera that flies through the glass, shrinks to millimeter scale and tours the inside of the machine, over the motherboard, through a canyon between heatsink fins, and under the graphics card fans. Same prompt, same model, two builds:
- AI Builder Only: Claude Fable 5.2 built the world and checked its own work. It ran 11 checks and called it done after 65 minutes.
- AI Builder with Critic Agents: the same model built the world, but every round a brand new Critic Agent reviewed screenshots of it and listed what looked wrong, ranked by severity. The builder fixed the list and the next critic looked again: 26 rounds, almost five hours.
Why does the Critic Agent work blind?
Each Critic Agent saw only the pictures. No code, no build notes, no memory of the round before, and no word from the builder about what it meant to do. That matters. When we replayed the real build's rounds, critics that read the builder's notes rated about 30% fewer problems as major than blind critics looking at the same screenshots. A reviewer who hears the reasons starts to accept them.
What did the Judge Agent find?
For the comparison, a separate Judge Agent reviewed screenshots of each finished world without being told which build it was looking at: Claude Fable 5.2, Claude Opus 5.5, and Claude Sonnet 5.5, three reviews each, nine per build. The share of reviews that rated each shot a major problem or worse:
- Heatsink canyon: 100% for AI Builder Only, 11% with Critic Agents
- Motherboard: 78% against 0%
- Graphics card fans: 67% against 0%
- Keyboard: 11% against 0%
- Wide desk shot: 11% against 89%
- Tower: 22% against 100%
Every model independently flagged the same glitch in the self-checked canyon, a spiky white-violet streak on the left wall, and described its walls as featureless teal fog. The room and the tower went the other way. Critic Agents also chase changes people barely notice: over 19 rounds the tower got darker, and the Judge Agent preferred the self-checked build's wide desk shot and tower. The Lab page shows all six shots, including the two the Critic Agent build lost.
Doesn't any reviewer catch obvious glitches?
Mostly, yes. In a control test we planted six glitches in copies of the world and asked reviewers to find them, 252 runs across Claude Fable 5.2, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. About 95% of planted glitches were caught whether the reviewer was blind, was the builder itself, or was the builder armed with excuses. Obvious problems get caught either way. The difference shows up in severity: what a reviewer decides is serious enough to fix, and when it decides the work is finished.
What can you do in the lab?
- Zoom in: six matching shots of both builds, side by side.
- Watch both tours: the full guided flight through each finished world.
- Fly it yourself: the Critic Agent build runs live in your browser on a computer with a dedicated graphics card.
- Get the free Critic Agent: paste-ready instructions, setup steps for Claude, ChatGPT, Copilot, Gemini or Grok, a guide to running the review loop, and a Claude skill. Free, no sign-up needed.
How did we build it?
Both worlds were written by Claude Fable 5.2 in Claude Code, from the same prompt, each as a single HTML file. The Critic Agent build ran 26 critic rounds, and the version in the comparison is round 26. A few later rounds added controls and faster first loading for the Lab page and are not part of the comparison. The tour videos were rendered frame by frame from each world's own guided tour, so neither build was recorded in real time.
What should you take from it?
An AI that checks its own work sees what it meant to build. When the result matters, add a second AI whose only job is finding problems, give it only the output and never the builder's explanation, have it rank what it finds, and fix the list before calling anything done. Expect it to be picky about things people may not notice, and keep a person in charge of what "finished" means.
Frequently asked questions
What is a Critic Agent?
A second AI, separate from the one doing the work, whose only job is reviewing the output and listing what looks wrong, ranked by severity. Ours started fresh every round and saw only screenshots.
Why keep the Critic Agent blind?
When we replayed the real build, critics that read the builder's notes rated about 30% fewer problems as major than blind critics looking at the same screenshots.
Did the Critic Agent build win every shot?
No. It won the close-ups: the canyon, the motherboard, the fans, and the keyboard. The Judge Agent preferred the self-checked build's wide desk shot and tower.
Does this work outside 3D?
The same loop works on anything a reviewer can look at, such as a report, a slide deck, a proposal or code. This test measured a 3D world, and we have not measured other kinds of work here.
Which AI tools does the free Critic Agent work with?
Claude, ChatGPT, Copilot, Gemini or Grok.
This build is a 3D world. The pattern shows up wherever AI does real work: it calls a report, a deck, or a proposal finished, and nobody looks closely until a client does. SquidTrain sets up Critic Agents in the AI tools your team already uses, and trains your team to run them. Book a free call.
Built with Claude Fable 5.2 in Claude Code. Judge Agent reviews by Claude Fable 5.2, Claude Opus 5.5, and Claude Sonnet 5.5.


