GPT agents beat Claude in Minecraft and Mindustry swarm tests

Y Combinator and Olam Labs tested groups of AI agents in Minecraft and Mindustry, as detailed in a research post. The experiments examined how agents form roles, communicate, and adapt while cooperating or competing.

In one Minecraft test, 50 GPT-6.1 Sol agents received three hours to build a civilization. They exchanged about 5,900 messages, followed a leader who emerged without a vote, after 46 of the other 49 agents first messaged the top name in their directory, and created rules for food distribution, construction, and resource management. After 22 deaths during the first night, the group reduced exploration and concentrated agents in enclosed shelters.

In nine 45 minute Bed Wars matches, teams of 10 GPT-6.1 Sol agents defeated teams of 10 Claude Opus 5.5 agents by six games to three. GPT agents sent between five and 30 times more messages, used compressed language, and focused on destroying the opposing bed. Claude agents played more defensively and often relied on kill totals.

GPT agents also won six of seven Mindustry matches against Claude agents. GPT teams divided work among more agents, changed tactical leadership, and downloaded the game's public source code to study its mechanics. Claude teams used proposal votes, retained the same leadership, and never searched the source code. The researchers released the Minecraft harness, the Mindustry environments, and the underlying multi agent framework as open source software.