AI agent swarms in Minecraft drifted into bureaucracy, while GPT teams beat Claude in tests
Reported by Hacker News and Tom's Hardware2 sources ↓
Researchers tested AI swarms in Minecraft survival and team games. A 50-agent colony became cautious and bureaucratic; GPT agents won a separate Mindustry series against Claude agents, according to Tom’s Hardware.
Key points
- The 50-agent Minecraft colony spent three hours trying to build a thriving civilization
- The agents sent about 5,900 messages and shifted toward shelter, food rationing and assigned tasks
- GPT agents beat Claude agents 6-1 in seven Mindustry games, Tom’s Hardware reports
- In Bed Wars, GPT agents played more aggressively while Claude agents favored defense, the researchers observed
Sources
Details
The researchers, working with Olam Labs, tested agent behavior in Minecraft and Mindustry. In the Minecraft survival test, 50 GPT-6.1 Sol agents were given three hours to build a thriving civilization. The researchers’ account says 22 of 34 agent deaths occurred on the first night, after which the agents became more risk-averse.
The colony developed food rationing, shelter rules and assigned operators rather than thriving. The researchers’ account says the agents sent about 5,900 messages; Tom’s Hardware highlights one episode in which four agents spent 11 minutes appointing an operator for a piece of fertilizer and a wheat block. The agents also created procedures after a cooked chicken was accidentally eaten by the wrong recipient.
The researchers also pitted 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents in Bed Wars. Tom’s Hardware reports that GPT agents devised more aggressive strategies and worked around a block-breaking bug, while several Claude agents concluded that breaking blocks was impossible. In seven Mindustry games, GPT agents won six, according to Tom’s Hardware.
Written from coverage as of