Explore tragedy-of-the-commons dynamics with AI coding agents

🐟 Fish Banks - Policy Sandbox

The simulation you played in the workshop, packaged so you can play with it and break it open with your AI coding agent. Same verified engine, no login, runs fully offline.

It's a tragedy-of-the-commons model: each fishing company chases cash; collectively they can overfish a shared ocean to collapse. The lesson isn't "be nice" - it's that the structure (a shared stock + a build delay + a "maximize my cash" goal) drives the collapse, and only the right rule prevents it.

The debrief reveal from the live workshop game: the hidden ocean (cyan) and catch-per-ship (orange) that no team could see while playing. Catch-per-ship was already sliding while cash still looked great - the early warning a real fishery actually gets. The single-player sandbox in this repo reproduces the same engine.

Want the detailed step-by-step (running locally, reading the charts, troubleshooting, extending)? See MANUAL.md.

Want the full model + multiplayer-architecture spec (parameters, equations, the Meadows leverage map, the 3 escapes, and how the live classroom game is built)? See FISHBANKS-SPEC.md.

Your cohort never collapsed it (room 8305: ocean survived, +$23.6M). But that was low fishing pressure + luck, not safety. You averaged ~1 ship bought/team/round - below the threshold. If any one team had floored it to buy-5, the shared ocean dies and everyone goes bankrupt. Your restraint was only as good as the least restrained player. This package lets you prove that to yourself in 10 seconds.

What a team saw on its own phone during the live game: its cash, its fleet, its last haul - but never the ocean stock. Players had to infer the shared commons from catch-per-ship alone. That hidden-state chokepoint is the whole lesson; the multiplayer architecture that enforces it is documented in FISHBANKS-SPEC.md Part 2.

Play it now (no setup)

Live (GitHub Pages): https://bayramannakov.github.io/fishbanks-sandbox/
Hit the ▶ Your cohort preset, then flip 😈 …now one of you defects.

Run it locally

Just open index.html - double-click it (or drag it into a browser tab). It's fully self-contained, so it works straight from your disk with no server and no install.

Explore headless (the fun part)

node explore.mjs

Prints the collapse threshold, your real game vs. one defector, and the enforced-quota cure - straight from the engine. Then go edit it.

The model in 30 seconds

Three functions, all pure, in fishbanks.js:

import { fishBanksParams, fbNewState, stepRound } from './fishbanks.js';

const P = fishBanksParams();        // tuned constants (K=6000, costs, regen curve)
let state = fbNewState(P, 4);       // 4 teams, 4800 fish to start
// one round = 3 simulated years; a decision per team:
const decisions = state.teams.map(() => ({ buy: 5, toSea: 99, sendNewToSea: true }));
const result = stepRound(state, decisions, P);   // mutates `state`, returns this round's outcome
// result.fish, result.catchPerShip, result.teams[i].roundCatch
// state.fish, state.teams[i].cash / shipsSea / shipsPort / bankrupt

A decision is { buy, toSea, sendNewToSea }:

• buy - ships ordered this round (they cost $25k and arrive next round - the build delay that causes overshoot).
• toSea - move ships from port to sea (99 = send them all). Ships at sea earn; ships in port idle cheaply.

Key numbers (engine-verified): carrying capacity 6,000, max sustainable yield ~8,688/round (regen peaks at 70% density), the commons cliff is ~75 ships at sea, a sea ship breaks even at 25 t/yr. With 3 teams it collapses at buy ≥ 3/team; with 4 teams at buy ≥ 3-4.

Explore with your agent (Claude Code / Codex)

This folder ships a CLAUDE.md / AGENTS.md so your agent already knows the engine. Open the folder in Claude Code or Codex and try:

Warm-ups

• "Run explore.mjs and explain why buy-2 survives but buy-3 collapses."
• "Find the exact MSY by sweeping a constant total catch per round."
• "Plot the fish stock for buy-4 as ASCII in the terminal."

Add a policy lever

• "Add supply-demand pricing: make fish sell for more as they get scarcer. Does collapse happen sooner or later? Why?" (it speeds up - scarcity rewards fishing harder)
• "Add a ship tax and a cap-and-trade quota. Which one sustains the fishery with the least lost profit?"
• "Add a subsidy on ship purchases and show it guarantees collapse."

Make it agentic

• "Replace the decision function with a call to an LLM - let Claude play one fishing company each round, given only what a real player sees (its own cash + catch, not the stock). Watch it strip-mine the ocean."
• "Now change only the goal you give the LLM from 'maximize my cash' to 'maximize total catch over 50 years without collapse.' Does it self-restrain? This is the whole point: the optimizer obeys the goal, not your intent."

Make it yours

• "Re-skin this as [your domain] - groundwater / API rate limits / ad-auction CAC / shared codebase tech-debt. Keep the same dynamics, rename the variables."

Files

From the AI + Systems Thinking course, Workshop 5 (Fish Banks). Model ported & tuned from Dennis Meadows' classic; engine validated round-by-round. Have fun breaking it.