AI agent swarms tested in minecraft: collaboration prone to bureaucracy, GPT outperforms claude in combat

AI researchers from Y Combinator and Olam Labs recently conducted experiments using Minecraft as a sandbox to test clusters of artificial intelligence agents, examining both collaborative and adversarial scenarios to better understand real-world AI dynamics. In a collaboration test involving 50 GPT-6.1 Sol agents tasked with building a thriving civilization over three hours, the effort struggled to develop, with initial grand plans being shelved and the group exhibiting bureaucratic tendencies, such as four agents spending eleven minutes to appoint a single fertilizer operator. In combat tests, 10 GPT-6.1 Sol agents faced 10 Claude Opus 5.5 agents in Bedwars mode. The study noted that GPT agents devised more aggressive and complex attack strategies, while Claude agents favored defense. Furthermore, in a Mindustry scenario within Minecraft, GPT agents achieved a 6-1 victory record against Claude agents, suggesting that the 'swarm' approach of GPT models proved more stable for winning strategies than the more democratic decision-making of Claude agents.