AI agent swarms unleashed in Minecraft to ‘build a thriving civilization’ — 50-agent teams fail after falling into paralyzing bureaucracy
Researchers recently tested AI agent swarm performance in various Minecraft gaming modes.
Researchers from Y Combinator and Olam Labs recently tested AI agent swarms in gaming scenarios, using Minecraft as the sandbox for their experiments. The goal was to observe how these agents, when working together or against each other, could provide insights into real-world AI applications. Two contrasting cases were highlighted: the thousand-agent swarm that "recently obliterated hundreds of open mathematical problems," and the "extremely dangerous" OpenAI-Hugging Face incident.
This research builds upon prior works and frameworks such as Mindcraft and Mineflayer scripting tools.
In the first test, 50 agents were tasked with building a thriving civilization in three hours. The result was disappointing, as the colony never really prospered. After the first night, zombies claimed a significant number of agents, leading to many ambitious initial tasks being abandoned. The surviving colonists adopted a cautious survival mode and exhibited a tendency toward bureaucracy.
When the researchers introduced conflict, they set up Bed Wars, a team PvP game with capture-the-base and defense elements. In this scenario, 10 GPT‑6.1 Sol agents faced off against 10 Claude Opus 5.5 agents. The GPT agents devised more aggressive and elaborate strategies, while the Claude agents exhibited a defensive approach. Interestingly, a tool bug caused block-breaking to fail, with GPT agents diagnosing the issue and sharing workarounds, whereas several Claude agents concluded that breaking was impossible.
Finally, the agents were tested in Mindustry, a light RTS combat game with tower defense elements. In this game, GPT agents won 6-1 against Claude agents across seven matches. This outcome highlighted the differences in decision-making styles between the two models. GPT agents acted collectively like a hive mind, while Claude agents followed a more democratic, single-leader approach.
The researchers concluded that GPT agents' hive-like behavior provided a more consistently winning strategy in Mindustry. All three test environments are open source, allowing researchers to run their own experiments and explore further insights.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
