Thousand Token Wood: shipping a multi-agent economy on a 3B model
A hackathon project used a 3B model to run a tiny woodland economy with five agents, emergent trade, and reskinned market shocks.
Intelligence analysis by GPT-5.4 Mini
Thousand Token Wood is a field report on making a real-time multi-agent economy work on Qwen2.5-3B. The project leans on engineered scarcity, prompt fixes, and driftable prices to turn a small model into a convincing trading sim.
It is like a tiny forest market where five animal shopkeepers trade food and wood to stay warm and fed. A small model runs each animal, and the creator had to add shortages and rules so the animals would actually buy, sell, panic, and make the town feel alive.
Analysis
What the project is
Thousand Token Wood is a tiny economy built for the Build Small Hackathon. Five woodland creatures, each controlled by Qwen2.5-3B, trade five goods for pebbles, gossip, hoard, and panic. The system runs on vLLM through Modal, with a Gradio app as the interface, and the author says the open traces show the full agent loop.
What made it work
The first version failed because the economy was too self-sufficient. Production outran consumption, so nobody had a reason to trade and the market went quiet. The fix was scarcity by design: each creature can only eat one unit of any single food per meal, some food spoils if hoarded, and every creature must burn firewood each turn. Because only one creature produces firewood, demand rises into a genuine bottleneck and the woodcutter becomes powerful.
What the small model could and could not do
The author reports that the model produced valid JSON on every call, which is useful for structured agent loops. But its judgment was weak at first: a creature would try to buy the very good it produced. Instead of switching to a larger model, the author tightened the prompt by explicitly stating what each agent makes, what it must never buy, and what it still needs. That improved decision quality. A tolerant parse-and-repair layer also prevents malformed output from crashing the sim; bad responses just become no-ops.
Making the world readable
The project also changed how wellbeing worked. A simple accumulator pushed creatures into a death spiral, so the author replaced it with a mean-reverting mood that recovers when a creature is fed and warm. Stakes are carried by pebbles, prices, and status, not by making the agents starve.
Why the demo feels alive
A key feature is the Wood Legend system, which reskins historical crises as folklore. For example, the Great Acorn Mania, the Hollow Log Trading Company, and the Run on Oona's Hoard each trigger real shocks that the agents react to. In one run, the rumor that Oona's vault was empty led her to sell honey, and the price fell from 10 to 3 over a few turns. The author says prices were initially frozen because agents anchored to a reference price, so the fix was to let prices drift with residual supply and demand after each round.
What the run showed
In a representative 15-turn run with a drought and a winter rumor, JSON stayed valid across all 75 calls, trade continued every turn, honey dropped from 10 to 3 during the bank-run legend, firewood rose from 4 to 7 under scarcity, and the wealth gap widened from 0.14 to 0.38. The woodcutter ended richest and the hoarder broke.
Key points
- The project uses five Qwen2.5-3B agents to simulate a woodland economy in real time.
- The first version failed until scarcity was engineered into food, spoilage, and firewood demand.
- The model reliably produced valid JSON, but needed stronger prompts to make sensible trade decisions.
- Historical market crashes were reskinned as woodland legends that triggered real shocks in the sim.
- A 15-turn run showed sustained trading, rising firewood prices, a honey crash, and a wider wealth gap.
If the design holds up, the project shows that small models can power real-time multi-agent worlds when the system is carefully shaped around them. The open traces and simple interface also make it easier for other builders to study and reuse the approach.
The article also shows that small models are fragile at the level of judgment, even when their output format is reliable. Without careful prompt design, scarcity mechanics, and price updates, the simulation can stall, anchor, or degrade into unrealistic behavior.
