Inspired by Tony Hawk's Pro Skater (Dreamcast) &
this X post of an FPS built by AI agents.
The test: how far do today's models go when they design, build, generate art, critique themselves & QA, with a human only steering?
The build
Six versions in one session. It started 2D when the first sandbox blocked 3D libraries, then went full Three.js once self-hosted. From there I handed 47 agents a loop: render the game, let a harsh critic panel score the frames, rebuild. An image model generated the entire cast & textures. A separate model QA'd it on a live mobile browser. Two taste passes landed the clean blocky look you're playing.
AI-made cast & textures. No human artist.
In-game: link combos, stomp tricks, a dusk skatepark.
Five models, matched to the job
| Opus 5 | Orchestrated the build & ran the multi-agent glow-up. The judgment. |
| Sonnet | Built the specified work: character select, collision, restyles. Cheap. |
| Codex / GPT | Ran real-browser QA & GitHub plumbing. $0 Claude quota. |
| Gemini (image) | Made every character & texture. |
| Fable 5 | Brokered the cross-model hand-offs. |
By the numbers
47
agents
~2.7M
tokens, that step
~9h
AI worked
5
AI models
What I'd take away
- "One prompt" is marketing. The leverage is orchestration: fan-out, loops, critic agents. In reality this took many iterations, not one shot.
- Models grade their own visuals now. Screenshot, harsh critic, rebuild. That loop is the unlock.
- Route by difficulty. Frontier judgment on one tier, cheap or free execution everywhere else.
- The walls are real. Matching a concept, physics feel & taste still need a human in the loop.