Thousand Token Wood: shipping a multi-agent economy on a 3B model

Source: HuggingFace Date: 2026-06-05 URL: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim

Summary

Thousand Token Wood is a real-time multi-agent economic simulation built for HuggingFace’s Build Small hackathon, running five woodland creature agents trading goods on a market — powered entirely by Qwen2.5-3B. The five agents manage scarcity (dietary variety, food spoilage, a winter fuel crisis) and respond to player-injected historical market scenarios like bank runs. Key findings: the 3B model generated valid JSON 100% of the time but made poor economic decisions without sharp prompting and computed constraints. Designed scarcity (telling agents what they produce and must never buy) proved essential — unconstrained agents produced a silent, uninteresting market. The project demonstrates that small models are fast and cheap enough for real-time multi-agent simulation where frontier models would be cost-prohibitive.

Implications

• Local models thread. 3B as the right-sized model for multi-agent simulation is a meaningful calibration point: not because 3B reasons well, but because it reasons well enough when constraints are externally computed, and it’s fast enough to run many agents per tick without cost becoming prohibitive. The pattern generalizes beyond games.
• Agentic engineering patterns thread. “Reliable formatting, unreliable reasoning” is the honest characterization of what small models contribute to multi-agent systems. The architectural lesson — compensate for weak reasoning with computed constraints and sharp prompting rather than hoping the model figures it out — is directly applicable to production agent system design. The economic simulation also surfaces emergent behavior without scripting, which is the goal for autonomous agent swarms.
• Coding agents thread (oblique). The hackathon context is signal in itself: developers are now comfortable building multi-agent systems on small local models as a first-class design choice, not a compromise. That normalization shifts what the baseline expectation is for “how many agents does this task need.”