What is Thousand Token Wood, and how is it different from a typical LLM multi-agent demo?

Thousand Token Wood is a multi-agent economic simulation built for the Build Small Hackathon, using the small Qwen2.5-3B model to drive five forest animals trading five goods for pebbles in a fictional market. The difference is that it pushes batched inference to the extreme — one GPU call per round handles every character's decision, and with vLLM deployed on Modal, the cost of running continuous simulations stays practical instead of just being a chat demo.

Why can a 3B small model produce bubbles and crashes without needing GPT-4-level reasoning?

The key isn't the model's reasoning power, it's the environmental constraints. The author added three scarcity rules — one unit of the same food per meal, food spoils and can't be stockpiled, and only one supplier of firewood in winter — and trading incentives emerge naturally. Over 15 rounds, honey dropped from 10 to 3, firewood rose from 4 to 7, and the Gini coefficient expanded from 0.14 to 0.38, all driven by the rules, with little connection to model size.

How did Qwen2.5-3B perform in this system? Was the output stable?

Across 75 API calls it hit 100% valid JSON output, so the format side was completely stable, with 3 to 9 trades closing per round. But its economic reasoning was weaker, leading to some irrational buy/sell behavior. The author's fix was to explicitly list each character's production goods, banned purchases, out-of-stock items, and examples in the prompt — shifting the judgment burden onto structured input instead of upgrading the model.

I'm building a multi-agent system — how should I pick a model? When should I switch to a bigger one?

Thousand Token Wood's answer is structure over scale. First check three things: does the environment have a scarcity mechanism that gives agents a reason to interact, does the prompt clearly list each character's state and banned items, and is the output format constrained by a JSON schema. If behavior still breaks down after tightening all three layers, then consider a bigger model. Most of the time it's a rule-design problem, not a parameter-count problem.

Roughly how much does it cost to run a setup like vLLM plus Modal plus Gradio yourself?

Thousand Token Wood does just one batched inference call per round covering every character, for 75 API calls across 15 rounds. Running Qwen2.5-3B on Modal's per-second-billed GPUs, a continuous simulation of several dozen rounds typically costs single-digit-dollar cents, far below GPT-4-level API pricing. Actual numbers vary with the Modal GPU type and cold-start time — it's worth running 5 rounds first to estimate.

What are the most common pitfalls when running a multi-agent economic simulation on a small model?

Three common mistakes: first, no scarcity design, leading to overproduction, vanishing trade incentives, and agents that don't interact; second, prompts that only give character names without listing what they can produce and what's banned to buy, causing the small model to buy randomly; third, expecting the model to reason out supply-and-demand logic on its own, which makes prices swing erratically. The fix in every case is to strengthen the external structure, not swap the model.

Who is this Thousand Token Wood approach a good fit for, and where doesn't it apply?

It's a good fit for researchers and independent developers doing agent simulations, game-theory experiments, teaching demos, or low-cost continuous inference. It's not suited for scenarios that need long-term memory continuity, complex natural-language dialogue, or real financial decision-making — a 3B model's economic reasoning is still weak, and it can only produce statistically reasonable behavior under tightly constrained rules, not something to plug directly into a production trading or advisory system.

Thousand Token Wood: Running a Multi-Agent Economy on a 3B Parameter Model

This article is a deep-dive from JudyAI Lab — an AI engineering playbook series with 100+ published guides, 5,000+ weekly readers across 60+ countries, focused on the practical side of running AI agents, trading systems, and content pipelines in production.

📰 Key Takeaways

Thousand Token Wood is a multi-agent economic simulation submitted to the Build Small Hackathon, using the small Qwen2.5-3B model to drive five forest animal characters trading five goods for pebble currency in a fictional market. The whole system runs on vLLM deployed on Modal, with a Gradio frontend, needing just one batched GPU call per round to resolve every character’s decision — keeping continuous simulation affordable.

The team found that without a deliberately designed scarcity mechanism, overproduction would wipe out any incentive to trade, so they added three constraints: only one unit of the same type of food per meal, food spoils and can’t be stockpiled, and firewood demand spikes in winter while only one character supplies it. These three rules directly produced bubbles and crashes — in a scenario modeled on the 1929 bank run, the character Oona dumps honey for pebbles, driving the honey price from 10 down to 3 within a few rounds; firewood, meanwhile, rose from 4 to 7 amid the winter crisis.

Across a 15-round test, 75 API calls achieved 100% valid JSON output, with 3 to 9 trades closing per round, and the Gini coefficient widened from 0.14 to 0.38 as wealth disparity emerged naturally. While the model’s JSON formatting was stable, its economic reasoning was weaker — the fix was to explicitly list each character’s production goods, banned purchases, out-of-stock items, and examples in the prompt, rather than switching to a bigger model. The author’s core conclusion: structure beats scale.

💬 JudyAI Lab Take

Thousand Token Wood produced bubbles and wealth divergence using a small Qwen2.5-3B model, and it points to something counterintuitive: you don’t need a bigger model, you need better rule design.

The system got honey to drop from 10 to 3 and firewood to rise from 4 to 7 within a handful of rounds — not through the model’s economic reasoning, but through three human-designed scarcity rules: food spoilage, a one-unit-per-meal cap, and only one supplier in winter. That gave the characters a real incentive to trade, and let the bubble emerge naturally. With each character’s production goods, banned purchases, and out-of-stock items spelled out explicitly in the prompt, 75 API calls hit 100% valid JSON output, and the Gini coefficient widened from 0.14 to 0.38 — wealth divergence showed up without anyone designing for it. The takeaway from this case: when a multi-agent system’s behavior falls short of expectations, tighten the environmental constraints and make the prompt more concrete first, before rushing to swap in a bigger model.

If you’re designing a multi-agent pipeline, try asking one question: strip away every external constraint — do the agents still have a reason to interact with each other? The answer usually lives in the rule design, not the model size.

📅 Original Article Info

Published: 2026-06-05T22:18
Source: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim

Thousand Token Wood: Running a Multi-Agent Economy on a 3B Parameter Model

📰 Key Takeaways

💬 JudyAI Lab Take

📅 Original Article Info

🔗 Further Reading

References

📰 Key Takeaways#

💬 JudyAI Lab Take#

📅 Original Article Info#

🔗 Further Reading#

References#

Get our weekly AI digest:

📰 Key Takeaways

💬 JudyAI Lab Take

📅 Original Article Info

🔗 Further Reading

References