Agents in Games Tests AI Swarms

The intriguing result is less “GPT beats Claude” than “coordination style matters as much as raw model capability.”

• –Fifty GPT agents formed leaders, rationing systems, and bureaucratic processes in a three-hour Minecraft colony.
• –GPT swarms communicated more aggressively and frequently; Claude swarms used slower, more democratic voting structures.
• –GPT agents proactively read Mindustry’s source code and shared tactical workarounds, while Claude agents often repeated documented mistakes.
• –The results are useful for multi-agent-system design, but the small sample and different harnesses make this an exploratory benchmark, not a definitive model ranking.
• –Multi-Agent Mindustry and Multi-Agent Harbor are open-sourced for developers to reproduce and extend the experiments.