A hackathon field report on building Thousand Token Wood, a multi-agent economic simulation where five AI agents (each running on Qwen2.5-3B) trade goods, gossip, and react to market shocks. Key engineering lessons include: small models reliably produce valid JSON but reason poorly without sharp prompting; emergent economic behavior requires designed scarcity (diet variety, spoilage, fuel crises); and prompt engineering outperforms scaling for improving decision quality. The simulation successfully reproduced historical market dynamics like bank runs and price crashes using reskinned 'Wood Legends,' with a Gini coefficient widening from 0.14 to 0.38 across a representative run. Served via vLLM on Modal with a Gradio frontend.
Table of contents
Why small is the design, not the limitThe first economy was dead on arrivalValid JSON, weak judgmentThen it started telling storiesWhat actually happenedTakeaways for building with small models4.8K Impressions