A hackathon field report on building a multi-agent simulation game where each AI agent runs on a different lab's small model (OpenAI gpt-oss-20b, MiniCPM3-4B, Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B). The post covers key engineering lessons: serving heterogeneous models via vLLM required a CUDA devel image fix; a tolerant JSON parse-and-repair layer made adding models trivial; secret information (insider tips) must be kept off-prompt and verified by automated tests to prevent leakage; and persistent agent memory should use bounded summaries rather than raw history to avoid prompt inflation. The result is a playable finance simulation where agents with genuinely different model behaviors create emergent market dynamics.
Table of contents
Heterogeneity is the product, not a constraintInformation asymmetry needs a firewallMemory is cheap drama if you bound itWhat actually happenedTakeaways for building with small models4 Impressions