# Thousand Token Wood: emergent market drama from 3-billion-parameter agents

*Field notes from the Build Small Hackathon, June 2026.*

The premise of the Build Small Hackathon is a quiet rebuke to the last two years of
AI: instead of wiring a giant model into yet another B2B tool, take 32 billion
parameters or fewer and build something delightful. I went smaller than the cap and
asked a narrower question. Can a 3-billion-parameter model, given a role and a
budget, produce the kind of emergent market behavior I spend my professional life
studying? The answer turned out to be yes, and the way it failed first was as
instructive as the way it eventually worked.

## Why small is the point, not the constraint

Thousand Token Wood is a tiny economy. Five woodland creatures, each its own agent,
trade five goods for pebbles, gossip, hoard, and panic. The player pokes the system
and watches bubbles, crashes, and a widening wealth gap appear on their own.

A living economy needs many agents thinking many times. That is precisely the regime
where a frontier model is the wrong tool: too slow and too expensive to run a council
of traders every tick. A small model is what makes a real-time multi-agent
simulation feasible at all. The constraint is the design.

The cast runs on Qwen2.5-3B-Instruct, served with vLLM on Modal, with a Gradio app
as the window onto the wood. Every creature decides in a single batched GPU call per
turn.

## The first economy was dead on arrival

The naive version did nothing. Production outran consumption, so every creature was
self-sufficient, hoarded its own crop, and never had a reason to trade. The market
cleared once and went silent.

The fix was to engineer scarcity, the way a real economy has it:

- **Diet variety.** A creature can eat only one unit of any single food per meal, so
  surviving means buying foods it does not grow.
- **Spoilage.** Perishable food rots if hoarded, which punishes stockpiling and forces
  surplus to be sold while it still has value.
- **A winter fuel crisis.** Every creature must burn firewood each turn, the need
  rises as winter deepens, and only one creature makes firewood.

That last mechanic is the engine of the drama. Demand for fuel outstrips one
supplier, so the woodcutter gets rich and everyone else competes for warmth.

## The model was a worse trader than it looked

With scarcity in place, a new problem surfaced, and it is the honest lesson of
working with small models. The 3B emitted perfectly valid JSON every time (100% across
hundreds of calls), but its economic judgment was mediocre: a creature that produced
acorns would post an order to buy acorns, the one thing it had in surplus.

Valid format, weak reasoning. The fix was not a bigger model; it was a sharper prompt.
I told each agent explicitly what it produced and must never buy, computed the list of
goods it was actually short on, and gave it one worked example of a sensible move.
Decision quality jumped, and the creatures began trading to their roles.

The second small-model lesson came from my wellbeing mechanic. I had modeled it as an
accumulator: shortfalls subtracted, full provisioning added one. Over fifteen turns,
any chronic shortfall ground every creature to zero, a death spiral that exposed the
agents' imperfect optimization and, worse, was no fun to watch. I reframed wellbeing
as a mean-reverting mood that recovers when a creature is fed and warm and never falls
to zero. The stakes moved into pebbles, prices, and status, where they belong in a
sandbox.

## Then it started telling stories

The feature I am most pleased with is the one that ties the project to my actual
field. The player can draw a "Wood Legend": a famous episode from market history,
reskinned as woodland folklore. Tulip Mania becomes the Great Acorn Mania. The South
Sea Bubble becomes the Hollow Log Trading Company. The 1929 bank runs become the Run
on Oona's Hoard.

These are not just flavor text. Each legend fires real shocks into the economy, and
the agents react. In one run I drew the Run on Oona's Hoard, the rumor that the owl's
vault was empty. Oona, reading the panic, began liquidating her honey to raise pebbles,
and the flood of supply crashed the honey price from 10 down to 3 over the next turns.
A reskinned bank run caused an agent to dump assets, which moved a market price. None
of that was scripted.

For the price action to be visible at all, I had to let prices move. They were frozen
because the agents quoted back the reference price I showed them. The fix was to let
the market reference drift with residual supply and demand after each round: heavy
unfilled buying pushes a price up, a glut pushes it down. Prices now trend during
scarcity and stay calm in balanced trade, which is exactly the behavior you want.

## What actually happened

A representative fifteen-turn run, with a drought and a winter rumor injected partway:

| Metric | Result |
|---|---|
| Valid JSON actions | 100% (75 of 75 calls) |
| Trades per turn | sustained 3 to 9, never silent |
| Honey price | crashed 10 to 3 during the bank-run legend |
| Firewood price | rose 4 to 7 as winter scarcity bit |
| Wealth gap (Gini) | widened 0.14 to 0.38 |
| Outcome | the woodcutter ended richest, the hoarder broke |

*Table 1. One playthrough of Thousand Token Wood on Qwen2.5-3B. The drama is emergent, not scripted.*

## What I took away

Three things. First, a small model is a reliable format generator and an unreliable
reasoner, and most of the engineering is closing that gap with structure and prompting
rather than scale. Second, emergent systems need designed scarcity; abundance is
boring. Third, the most compelling demos sit at the intersection of a technical
constraint and something you already understand deeply. I did not have to invent the
drama. Three centuries of market history had it ready, and a council of 3B agents was
enough to play it out.

Small models, big adventures.

---

*Try it: the [Space](https://huggingface.co/spaces/build-small-hackathon/thousand-token-wood-sim).
Open agent traces: the [dataset](https://huggingface.co/datasets/build-small-hackathon/thousand-token-wood-traces).*