Introduction
AI researchers recently tested AI agent swarms in gaming scenarios. Minecraft and its various game modes were the sandboxes where hundreds of agents were unleashed to work together or as adversaries. The idea seems to be that the in-game collaborations and conflicts would provide insight into real-world cases with AI agents involved.
This study might help us see what happened, good and bad, in recent AI swarm cases, and thus developers of this technology can improve it. The research is built on prior work and frameworks, specifically repos like Mindcraft and scripting tools in Mineflayer.
Collaborative Civilization Building
The first AI swarm testing plan was a simple collaborative one. During this test phase, the researchers tasked 50 agents with building a thriving civilization in three hours.
Probably the most interesting insight from this test was that the colony never really thrived. After the first night saw zombies take out a large number of agents, it was noted that many ambitious initial tasks were abandoned.
From then on, the colonists cautiously survived, and there was a tendency toward bureaucracy. As an example, four agents spent 11 minutes appointing a sole fertilizer operator for a single piece of bone meal on a single wheat block.
Conflict Scenarios
Next up, the researchers decided to put AI agent swarms into conflict. First they tested them in a team PvP capture the base mode using 10 GPT-6.1 Sol versus 10 Claude Opus 5.5 agents.
GPT agents came up with much more aggressive and elaborate strategies, while Claude was more defensive. A tool bug stopped block-breaking, and GPT diagnosed it and shared workarounds while several Claudes concluded breaking was impossible.
Combat Results
Finally, the agents were tested in a light RTS combat game with tower defense mode in Minecraft. The GPTs won six to one against the Claudes, which was a strong result for the OpenAI-created agents.
Future Research Directions
These swarms are open sourced so you can run your own if you are interested in tweaking and testing. Suggestions for further study include comparing different mixes of models and varying sizes of agent swarms.
