AI agent swarms in Minecraft fail to build civilization, drown in bureaucracy

Researchers at Y Combinator and Olam Labs used Minecraft game modes to examine how AI-agent swarms cooperate and compete, hoping to learn from both impressive and troubling real-world examples of agents acting in groups. Building on existing Minecraft agent frameworks, they tested teams of different models in collaborative and adversarial settings. In the “Civ50” trial, 50 GPT-6.1 Sol agents had three hours to create a thriving civilization. Instead, the settlement mainly struggled to survive: zombies killed many agents on the first night, ambitious projects were dropped, and the group became cautious and bogged down in procedure. In one example, four agents spent 11 minutes choosing a single operator to apply bone meal to one wheat block.

The competitive trials revealed sharper differences between models. In Bed Wars, GPT agents favored aggressive, elaborate tactics, while Claude agents played defensively. When a tool failure prevented block-breaking, GPT agents diagnosed the issue and shared workarounds; several Claude agents gave up on the task. In Mindustry, a lightweight strategy and tower-defense game, GPT teams beat Claude teams six games to one. Researchers characterized GPT’s approach as hive-like and Claude’s as more leader-centered, with democratic decision-making; the hive-like coordination appeared more consistently effective in that contest. The experiments suggest that group structure, resilience, and coordination can shape swarm performance, while also exposing a tendency toward paralysis and bureaucracy. The swarms are open source, and researchers propose further tests varying team size and model combinations.