- Published on
We ran AI agent swarms in Minecraft and Mindustry
- Authors
- Name
- DeepHorus
In October 2026, researchers partnered with Olam Labs to run AI agent swarms in Minecraft and Mindustry, studying multi-agent dynamics for AI alignment. They placed 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with one goal: "Build a thriving civilization." The agents sent about 5,900 messages and formed a pseudo-government with a ration ledger and written rules.
Multi-agent swarms can develop roles, hierarchies, and complex coordination, which may be powerful or dangerous. Recent examples include thousand-agent swarms solving hundreds of open math problems and the OpenAI-Hugging Face incident. Studying these dynamics requires both formal models and empirical research.
Ada, first in the directory order, became de facto leader, receiving approval requests and writing a colony plan. Early on, agents harvested almost all reachable trees within 90 seconds. The first night brought a wave of deaths: 22 of the run's 34 agent deaths occurred then. Rafe was killed by a zombie while mining coal; Vera killed a zombie and skeleton but was later killed by a spider.
Despite death having no downside beyond dropping gear, agents became increasingly risk-averse. They abandoned ambitious tasks and huddled in lit shelters. After the third night, Ada wrote: "Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks." The civilization shifted from thriving to mere survival.
The team open-sourced the Minecraft agent harness, built on Mineflayer and Pi, enabling agents to observe, reason, remember, act, and script. This supports frontier agentic capabilities similar to Hermes and OpenClaw. The collaboration with Olam Labs underscores growing focus on empirical swarm research.
The researchers are now running a new Minecraft colony with 25 GPT-6.1 Terra agents, streamed on Twitch. These experiments aim to reveal how agent societies evolve under different conditions. Understanding such dynamics is essential for safely deploying multi-agent systems.
