Imagine a bustling woodland economy where five charming creatures, each a unique agent, trade, hoard, gossip, and even panic. This isn’t just a whimsical tale; it’s Thousand Token Wood, a multi-agent simulation born from the recent Build Small Hackathon. We’re talking about a tiny, self-contained market, powered by a 3-billion-parameter language model that brings its denizens to life.
This report delves into the engineering marvel behind Thousand Token Wood, offering a candid look at what a 3B model can and cannot achieve in a dynamic economic system. We’ll explore the ingenious solutions that transformed a silent, static market into a vibrant, ever-changing drama of bubbles, crashes, and widening wealth gaps. If you’re building with smaller AI models, this is a field report you won’t want to miss.
Why Small Models Are a Big Deal for Multi-Agent Systems
When creating a living economy with numerous agents making decisions in real-time, the choice of model size becomes critical. Frontier models, while powerful, are often too slow and prohibitively expensive to run a council of traders repeatedly every turn. This is precisely where a small model shines, making a real-time, multi-agent simulation not just possible, but practical.
In Thousand Token Wood, every creature makes its economic decisions in a single, batched GPU call per turn, enabling a truly responsive and interactive experience. This efficiency allows the simulation to run smoothly, showcasing the incredible potential of optimizing model size for specific, high-frequency tasks. It’s a testament to how “small” isn’t a limitation, but a fundamental design choice for scalability and performance.
Engineering an Economy: Scarcity and Smarter Agents
The initial version of our woodland economy was, frankly, a bit dull. Production consistently outpaced consumption, leading to self-sufficiency among the creatures and a market that cleared once before falling silent. The crucial lesson here was that emergent systems, especially economies, thrive on designed scarcity.
To breathe life into the wood, several key mechanics were introduced. Production became variable, tied to a die roll, while consumption was linked directly to each agent’s wellbeing. Furthermore, needs were engineered to grow over time, ensuring a constant demand, and certain goods, like warmth from the woodcutter, had only one supplier.
This strategic scarcity fundamentally changed the dynamics, creating compelling drama. When a single supplier cannot meet rising demand, the woodcutter flourishes, while everyone else fiercely competes for essential resources. This imbalance sparks the very trading behavior we wanted to see.
With scarcity addressed, another challenge emerged: the 3B model, while consistently emitting valid JSON responses on 100% of calls, exhibited surprisingly poor economic judgment. For instance, a creature producing acorns might paradoxically post an order to buy acorns—the very item it had in surplus.
The solution wasn’t a larger model, but a sharper, more refined prompt. Each agent was explicitly informed of what it produced and should never buy. We also computed the exact list of goods each creature genuinely needed and provided a clear, worked example of how to make a trading decision. This dramatically improved decision quality, allowing agents to trade intelligently and according to their roles.
To ensure robustness, the entire loop is wrapped in a tolerant JSON parse-and-repair layer. This clever addition prevents the simulation from crashing due to a malformed response, instead gracefully degrading it into a no-op. We also refined the concept of “wellbeing,” reframing it from a fragile accumulator that could lead to death spirals into a mean-reverting mood. This ensures creatures recover when fed and warm, never hitting zero, keeping the focus on economic stakes rather than existential dread.
Bringing History to Life: Market Shocks and Storytelling
One of the most engaging features of Thousand Token Wood is its ability to draw a “Wood Legend.” Players can select famous episodes from market history, which are then charmingly reskinned as woodland folklore. Imagine the Tulip Mania becoming the Great Acorn Mania, or the 1929 bank runs transforming into the Run on Oona’s Hoard.
These legends are far more than mere flavor text; they trigger real, dynamic economic shocks that the agents react to unscripted. For example, selecting the Run on Oona’s Hoard rumor (that the owl’s vault was empty) caused Oona to liquidate her honey to raise pebbles. This flood of supply crashed the honey price from 10 to 3 over subsequent turns—a true market reaction, driven by an agent’s fear and the emergent system.
For such emergent behavior to be visible, prices themselves needed to be dynamic. Initially, they remained static because agents simply quoted the reference price provided to them. The fix was elegant: allowing the market reference to drift with residual supply and demand after each round. Heavy, unfilled buying now pushes prices up, while a glut drives them down, enabling realistic price trends during scarcity and calm stability during balanced trade.
Key Takeaways for Small Model Builders
Building Thousand Token Wood offered profound insights into working with smaller language models. Much of the engineering effort centered on bridging the gap between a small model’s reliable formatting capabilities and its often-unreliable reasoning. This was achieved through meticulous structural design and precise prompting, proving that sophistication can come from clever design rather than sheer scale.
Furthermore, the project underscored that emergent systems absolutely require designed scarcity. Abundance, while seemingly desirable, makes for a rather boring simulation. And perhaps most importantly, the most compelling small-model demos don’t need fabricated drama. Leveraging centuries of real market history provided a rich tapestry of scenarios, which a council of 3B agents proved more than capable of playing out.
Thousand Token Wood demonstrates that small models aren’t just for niche applications; they’re capable of enabling big adventures and complex, emergent behaviors. It’s an exciting time to explore the vast possibilities that optimized, efficient AI models unlock. Try the Space to experience this tiny economy yourself!
Source: Hugging Face Blog
