WYC experiment in Minecraft sees fifty intelligent agents shift from building a civilization to regulating food and avoiding hazards

The WYC labs, in collaboration with Ulum Labs, launched a live experiment to study the dynamics of artificial intelligence swarms by deploying fifty agents powered by the GPT-6.1 Sol model in the Minecraft game world for three consecutive hours, with a single explicit goal: building a thriving civilization. The researchers equipped the models with an agent guide, a shared execution space, and the ability to exchange direct messages through an open-source framework based on Mineflayer tools and the Bay runtime loop, aiming to test swarm behavior when faced with an open environment that imposes survival challenges, resource gathering, and shelter construction.

The first minutes revealed a consumption rush that quickly produced a spontaneous leadership hierarchy.In just ninety seconds from start, the models depleted all trees around the starting point, after which an administrative hierarchy based entirely on the alphabetical order in the directory emerged. Because it topped the list, the agent Ada assumed actual leadership, and the remaining agents messaged it for approvals; it placed the settlement plan in the shared space, and allocated tasks among storage warehouse management, wheat farm harvesting, stone shelter provisioning, and prohibiting local wood depletion.

However, the first night’s solutions in the game caused a fundamental shift in swarm behavior, as virtual monster attacks killed 22 agents before the twentieth minute, out of a total of 34 deaths during the experiment. Although death in the test environment incurred no cost beyond resurrection, loss of equipment, and a note in the death log, the models became overly cautious; they abandoned large expansion tasks, clustered in enclosed shelters, and Ada issued strict orders to halt exploratory digging and remain stationary in illuminated fortifications until short, safe tasks could be assigned.

The civilizational ambition gradually turned into a bureaucracy occupied with survival and ration distribution.With the models exchanging roughly 5,900 messages, activity centered on a food-ration ledger managed by the agent Ines, where each meal required extensive verification and matching to energy status before allocation. Operational coordination failures quickly emerged, including delayed assistance to isolated agents, suspension of resource extraction until the hungry were fed, and repeated inquiries about ration delivery without resolution, causing the entire computational effort to shift to managing internal constraints rather than expanding.

These results reveal to technology managers and systems engineering teams in the Gulf, Egypt, and the region the limits of betting on independent swarms; a rush to delegate smart swarms to manage supply chains or banking operations without prior coordination engineering leads to depletion of inference budgets in sterile bureaucratic correspondence. If you are building a multi-agent architecture for your organization, the greatest risk does not lie in a single model’s failure, but in its intrinsic tendency to contract and favor defensive actions when encountering the first operational error, generating random hierarchies that disrupt workflow unless governance and recovery protocols are pre-configured with decisive programmatic rules.