AI Agent Lab: Critic
Use an AI Critic Agent to Fix the Problems the AI Builder Missed
We built two versions of an interactive 3D world using the same prompt and the same model (Claude Fable 5.2). The goal was a late-night trading desk you can fly into and tour, all the way down to the computer's motherboard and graphics card. One AI Builder just checked its own work and called it done in about an hour. We asked the other AI Builder to use new, blind AI Critic Agents to review every round without any prior knowledge, and after 26 rounds over almost five hours, we were amazed at the difference!
Compare two "finished" 3D worlds built by:AI Builder Only vs. AI Builder with Critic Agents
- Pick a shot
- Drag over either picture (both magnifiers move together)
- See what the Judge Agent found
Six shots from each final 3D world, scored by the Judge Agent
We asked a new Critic Agent to judge both finished builds. We call it the Judge Agent. It was never told which build it was looking at, and three AI models each ran three independent reviews of each build. The wide shots of the office favored the AI Builder Only version. Inside the computer, where the fine detail lives, the AI Builder with Critic Agents won in every shot.
How the Critic Agent loop works
Two agents with two jobs. The AI Builder writes and fixes. The Critic Agent only looks, and a brand new Critic Agent starts every round.
- The AI Builder renders
Claude Code builds the world and renders six fixed shots, close-ups and frames from the tour.
- A new Critic Agent looks
It sees only the pictures. No code, no notes, no earlier reviews.
- It ranks what's wrong
Every problem gets a label (Critical, Major, Moderate or Minor) and a place in the frame.
- The AI Builder fixes, repeat
26 rounds, until what was left were judgment calls for a person to make.
Why the Critic Agents work blind
We replayed 23 rounds of the build with two kinds of Critic Agent looking at the same pictures. One saw only the pictures. The other also read the AI Builder's notes on what it had just fixed.
The Critic Agent that read the notes listed more problems overall (about 19 per review against 13), yet rated fewer of them major. Once it knew the AI Builder had "fixed" something, it went easy on it, so a loop run that way would likely have called the world done much sooner.
See both final worlds in motion
Checking your graphics card…
What this means for your AI work
It stopped after 11 self-checks. From across the room the result looked finished. Up close it wasn't.
With no backstory and only the result to judge, Critic Agents kept finding real problems for 26 rounds.
Critic Agents also push changes a person might not want. Over 26 rounds the tower view ended up darker than in earlier rounds, and the Judge Agent rated a major problem in it in every review. Let AI find the problems. You decide what good looks like.
Let us build Critic Agents into your team's AI work
An AI Builder with Critic Agents works for reports, proposals, spreadsheets, code and designs, not just 3D worlds. We set up the workflow in the AI tools your team already uses, then teach your team to run it themselves.
The Critic Agent
Free, no sign-up needed. Instructions for a Critic Agent that judges only the finished work and ranks what's wrong, plus setup notes for Claude, ChatGPT, Copilot, Gemini or Grok.
Download the Critic Agent (zip)
Find light in the deep waters of AI
Let's talk! A short call will help you learn more about our training and consulting solutions. We can also discuss whether our existing licensed tools can help, or whether you need a custom build.