Researchers test AI agent swarms in Minecraft: collaboration tends toward bureaucracy, GPT has a higher win rate against Claude

Researchers Test AI Agent Swarms in Minecraft: Collaboration Tends Toward Bureaucratization, Higher Win Rate Against GPT

English text is being prepared.

English text is being prepared.

English text is being prepared.

Multi-agent interaction in Minecraft

The first agent swarm test was a simple collaborative task. In this test phase, the researchers arranged 50 agents and required them to build a thriving civilized settlement within three hours. This test used 50 GPT-6.1 Sol agents and did not call on any Claude-series agents.

The most noteworthy conclusion of this test, called Civ50, is perhaps that this settlement of 50 agents never truly grew and prospered. On the first night, a large number of agents were killed by zombies (though the penalty mechanism had only a mild effect), and the research records show that "many of the grand tasks set at the beginning were shelved."

After that, these colonists could only carefully maintain survival, and there also emerged a tendency toward bureaucratization-like behavior. The article gives an example: "Four agents spent 11 minutes specifically appointing a 'sole fertilizer operator,' while the task was merely to use one bone meal on a wheat field." The researchers said that the behavior patterns of this kind of agent swarm are very similar to phenomena they observed in coding agent projects.

Next, the researchers arranged for the agent swarm to conduct confrontation tests. First, they chose the Bed Wars mode, a team PvP gameplay whose core is to destroy the opponent's base and defend one's own base. The test groups were 10 GPT-6.1 Sol agents versus 10 Claude Opus 5.5 agents.

The researchers recorded: "The offensive strategies conceived by the GPT agents were more aggressive and complex, while the Claude agents leaned toward defense." The differences in how the two teams handled problems were also interesting. "The block-breaking function had a tool malfunction; the GPT agents completed problem diagnosis and shared temporary workarounds; multiple Claude agents directly determined that the breaking operation could not be executed."

Finally, the research team conducted tests in Mindustry, a lightweight real-time strategy mod featuring tower defense gameplay, running within Minecraft. In seven matches, the GPT agents defeated the Claude agents six to one, a good result for this agent developed by OpenAI. The researchers concluded that the GPT agents were more like a unified swarm, while the Claude agents had a single leader and adopted a more democratic decision-making process. In the Mindustry scenario, the swarm model more easily formed stable winning strategies.

This agent swarm project has been open-sourced; if interested, you can adjust parameters and reproduce the tests yourself. The end of the original blog post lists some open questions for hands-on exploration and also gives directions for follow-up research. For example, different model combinations can be paired in Minecraft and the agent swarm scale can be adjusted, which is expected to yield valuable conclusions.

Reference materials:

English text is being prepared.