AI researchers at Ycombinator and Olam Labs tested AI agent swarms in Minecraft, using the game as a sandbox to study multi-agent dynamics. In one collaborative test, 50 agents, reportedly running GPT-6.1 Sol, were tasked with building a thriving civilization over three hours. The researchers noted that the colony struggled to thrive, observing that "many ambitious initial tasks were abandoned." A specific instance cited was four agents spending 11 minutes appointing a "sole fertilizer operator" for a single piece of bone meal on a single wheat block, a behavior the researchers stated "felt similar to patterns seen while working with coding agents."
In a conflict test within Bed Wars, the setup pitted 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents. The researchers observed that "GPT agents came up with much more aggressive and elaborate strategies," while the Claude agents tended to be more defensive. When a tool bug prevented block-breaking, the GPT agents diagnosed the issue and shared workarounds, whereas several Claude agents concluded that breaking was impossible.
Testing continued in Mindustry, a light RTS combat game mode within Minecraft. In seven games, the GPT agents reportedly won 6-1 against the Claude agents. The researchers concluded that the GPT agents acted more like a hive, which formed the basis for a more consistently winning strategy, while the Claude agents exhibited a single leader but a democratic-style decision process. The research, which builds on frameworks like Mindcraft and Mineflayer, is open sourced for further testing.
