Thousand Token Wood and the case for designing around small models
A Build Small Hackathon project ran a five-agent economy on Qwen2.5-3B and found that scarcity, prompting, and parse-repair did the work people usually assign to bigger models.
Builders, integrators, prompt engineers · 2 min read
Lester Leong's Build Small Hackathon writeup, Thousand Token Wood, is the kind of small-model field report worth reading closely. The setup: five woodland creatures, each an agent backed by Qwen2.5-3B, served via vLLM on Modal, with a Gradio front end. Every creature decides in a single batched GPU call per turn. That batching is the whole point — a frontier model would be too slow and too expensive to run a council of traders every tick.
The honest part of the post is where the 3B model fails. According to Leong, the model hit valid JSON on 100% of calls (75 of 75 in a representative run), but its economic judgment was poor: a creature producing acorns would post a buy order for acorns. The fix wasn't a larger model. It was a sharper prompt that told each agent what it produced and must never buy, computed the exact list of goods it was short on, and gave it one worked example. A tolerant JSON parse-and-repair layer turns malformed responses into no-ops instead of crashes.
The other engineering lesson is about system design, not the model. The naive economy did nothing — production outran consumption, so nobody had to trade. Leong engineered scarcity in three places:
- Diet variety: only one unit of any single food per meal.
- Spoilage: perishables rot if hoarded.
- Winter fuel: every creature must burn firewood each turn, demand rises, and only one creature produces it.
He also reframed wellbeing from an accumulator (which caused death spirals) to a mean-reverting mood that never hits zero. Prices, frozen at first because agents parroted the reference price, were unstuck by letting the market reference drift with residual supply and demand.
The pattern here is one builders should internalize: with a small model, most of the work is closing the gap between reliable formatting and unreliable reasoning — using structure, scarcity, and prompt scaffolding instead of parameters. The open traces dataset (full prompts, raw responses, parsed actions, private thoughts) is the kind of artifact more agent demos should ship.
If you're building multi-agent systems and reaching for a frontier model by default, this is a useful counterexample to keep on the shelf.
