We ran AI agent swarms in Minecraft and Mindustry

The project explores whether groups of AI agents can work together to solve tasks in complex games. It tests agent swarms in Minecraft, where players gather resources, craft tools, and build in an open world, and Mindustry, where they must organize production, construct defenses, and withstand enemy waves. These environments offer different tests of cooperation: Minecraft demands flexible exploration and planning, while Mindustry requires tightly coordinated, ongoing management of resources and threats.

The experiments highlight a gap between an agent’s ability to handle an individual action and a group’s ability to sustain a shared plan. Multiple agents can divide work and make progress, but coordination becomes difficult when goals are unclear, communication is inconsistent, or plans need to adapt over time. The games make those weaknesses visible because success depends on linking many steps together, not just choosing a plausible next move. The central takeaway is that adding agents does not automatically produce effective teamwork: swarms need clear roles, shared context, and reliable coordination to turn individual capabilities into dependable collective behavior.