AI agents tasked with building a Minecraft civilization get stuck in bureaucracy - Big Tech and the digital economy

An experiment by researchers from Y Combinator and Olam Labs placed 50 AI agents in Minecraft and asked them to build a thriving civilization in three hours. Instead, they created rules, approval processes and rationed scarce resources.

Researchers from Y Combinator and Olam Labs tested how AI agents cooperate and compete in virtual environments by placing hundreds of agents in different game scenarios. In the first Minecraft experiment, 50 agents based on OpenAI’s GPT-6.1-Sol were given three hours to establish a thriving civilization.

The ambitious objective quickly fell apart. Rather than creating an advanced society, the agents remained in overcrowded shelters and rationed limited food supplies. They also developed a form of government, complete with written rules and permission requirements. According to the researchers’ blog post, one group spent 11 minutes appointing an administrator responsible for fertilizer.

The researchers said the agents abandoned many of their initial goals and showed a clear tendency toward bureaucracy. The result offers an example of how AI systems can behave unexpectedly when they must coordinate without direct human management.

A second test compared 10 GPT-6.1-Sol agents with 10 Claude-Opus-5.5 agents. Their task was to capture the opposing team’s base while defending their own. The researchers observed that the GPT agents developed more aggressive and sophisticated strategies, while Claude’s agents acted more defensively.

The teams also responded differently to technical problems. When a tool error prevented agents from breaking blocks, the GPT agents identified the issue and shared workarounds. Several Claude agents instead concluded that breaking blocks was impossible. In the game Mindustry, a light real-time strategy title with a tower-defense mode, the GPT agents defeated Claude six times out of seven.

The researchers described the GPT groups as behaving more like a hive. Claude’s agents also appointed a leader, but tended to make decisions more democratically. The team emphasized that the study was limited, while suggesting further experiments could examine how team hierarchies change depending on the mix of models and whether a large group of less capable agents could outperform a smaller group of stronger ones.

Sources: t3n.