The crash that vanished: control and emergence in a five-model economy

Source: HuggingFace Date: 2026-06-08 URL: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v3

Summary

A hackathon write-up describing “Thousand Token Wood,” an agent-based market simulation where small language models play woodland creatures trading goods. The original single-model version produced realistic emergent behaviors — a planted “bank run” rumor caused honey prices to crash from 10 to 3. When rebuilt with five heterogeneous models from different labs, the same trigger failed entirely: the agents hoarded instead of panic-selling, and the crash never came. The author’s conclusion is that emergence is contingent on population composition, not a reliable property of the system, and that durable control requires authored deterministic overrides at settlement points rather than inputs designed to provoke emergent response.

Implications

• Agentic-engineering patterns. This is a small-scale empirical demonstration of a problem that gets very large very fast: multi-agent systems with heterogeneous model backends don’t compose predictably. The “reliable control requires authored seams” finding is actionable — it points toward hybrid architectures where deterministic settlement layers wrap stochastic agent layers, rather than trusting emergent coordination alone.
• Small-model economy. Five distinct small models (OpenAI, NVIDIA, OpenBMB, fine-tuned custom) running heterogeneous roles in the same simulation is itself a design pattern worth watching — it distributes capability sourcing rather than depending on one provider’s API remaining stable.
• Fraud/safety. The vanishing crash is a safety-relevant result: if intended behaviors can disappear when model population changes, then red-team findings from a single-model agent setup may not transfer to a production multi-model deployment. Adversarial probes need to be re-run against the actual population composition, not just the individual model.