Aquarium Burst — One Prompt, Full Physics

I gave Qwen 3.6 27B a single prompt describing a glass aquarium that cracks and bursts. What came back was a complete 3D simulation with water dynamics, buoyancy, and glass shatter — built by a multi-agent pipeline running entirely on one RTX 3090.

The Prompt

Create a large glass aquarium whose side panel develops a visible crack and then bursts. Water escapes through the opening with flow strength based on water depth and decreasing as the tank drains. A curved water jet affected by gravity. A spreading puddle that collides with the room boundaries. Fish, rocks, plants, and a floating toy reacting differently according to density, buoyancy, drag, and current.

That's it. No architecture doc, no tech stack specified, no wireframes. One paragraph describing the behaviour I wanted to see.


The Simulation

Drag the crack line up or down to set the burst point, then release. The water jet strength follows Torricelli's law — deeper water means a stronger jet. As the tank drains, the flow weakens and the waterline drops.


What It Does

The simulation implements 13 distinct physics behaviours:


The Numbers

906
Lines of code
13
Physics behaviours
1
Prompt
0
API costs

Built with Three.js r185 loaded from CDN. Single HTML file. No build step, no dependencies to install. Runs in any modern browser.


How It Was Built

This wasn't a single model call. The PI Agent Harness ran a 9-agent pipeline — each agent handling one job, coordinating through structured prompts and a shared task board. All running locally on a single RTX 3090.

The Pipeline

Researcher Researcher Architect Planner
Debugger Tester Debugger Tester Reviewer

What Each Agent Did

  1. Researcher (physics) — gathered the equations: Torricelli's law, drag coefficients, buoyancy, angular velocity for glass fragments
  2. Researcher (Three.js) — evaluated whether 3D rendering added value over Canvas 2D, recommended Three.js from CDN
  3. Architect — designed the system: scene graph, physics loop, entity classes, interaction model (30KB architecture doc)
  4. Planner — broke the build into 16 implementation steps with acceptance criteria
  5. Debugger — wrote the full 906-line implementation
  6. Tester — ran 14 tests: 14 PASS, 3 PARTIAL (fixable issues identified)
  7. Debugger (fixes) — applied 3 surgical fixes for the partial failures
  8. Tester (regression) — re-ran all 17 tests: 17 PASS, zero regressions
  9. Reviewer — final audit: 13/13 requirements verified, verdict: complete

Why It Matters

This is what a local AI agent team can ship. No API keys, no cloud credits, no vendor lock-in. One prompt describing what I wanted to see — and a pipeline of specialised agents that researched, designed, built, tested, and reviewed the whole thing on a single consumer GPU.

The harness that made this possible is documented in the PI Agent Harness bundle — ebook, video series, and all 13 agent definitions.