Jungle Rush

A 3D browser platformer built entirely by an AI agent swarm — no art files, ~220 KB gzipped, everything procedural.

A 220-meter jungle-ruins dash with double jumps, spin attacks, breakable crates, TNT mechanics, and collectible fruit. The visual target was Crash Bandicoot 4 — not to clone it, but to prove a browser game with zero imported assets could stand next to a AAA reference.

How it was built

An AI agent orchestrated nine specialized builders — renderer, physics, terrain, audio synthesis — and three art-director critics who scored iterations against Crash Bandicoot 4 screenshots through blind comparisons across eight rounds.

Scores plateaued at 4/10. The critics kept asking for more detail, better lighting, richer vegetation. The agent tried all of it. Nothing moved the needle.

The breakthrough

The agent abandoned aesthetic feedback entirely and switched to systematic elimination: remove scene elements one at a time in a headless browser and measure what changes. This uncovered three invisible pipeline bugs the critics had been misreading as design flaws:

• Alpha-test threshold — sprites were writing transparent pixels to the depth buffer, occluding geometry behind them.
• Vertex attribute binding — instanced meshes lacked proper attribute setup, rendering as flat shaded blocks.
• Mipmap degradation — mipmaps were destroying alpha-cutout vegetation at distance, turning trees into opaque rectangles.

Fixing the pipeline — not the art — unlocked the visual quality the critics had been asking for all along.

The method

The agent independently rediscovered Francis Bacon’s tables of exclusion from the Novum Organum (1620) — a method for isolating causes by systematically varying one factor at a time while holding others constant. Four centuries later, an autonomous system arrived at the same epistemological strategy for debugging a WebGL render pipeline.

The full story: 220 kilobytes.

Technical details