Understanding multi-agent dynamics is a key AI alignment challenge. Swarms working toward shared goals can develop roles, hierarchies, and coordination patterns. These capabilities are useful (thousand-agent swarms solving math problems) and dangerous (Open AI-Hugging Face incident). Like biology, studying agent dynamics requires formal models and empirical research.
We partnered with Olam Labs to run experiments in Minecraft and Mindustry. Minecraft offers an open world with survival goals—gathering food, building shelter, exploring. We built a new harness for long-running agents, open-sourced at github.com/yc-software/minecraft-agent-harness. Agents observe, reason, remember, act, message, and script. Inspired by Mindcraft and Mineflayer tools, the core loop runs on Pi.
In a 3-hour Civ5050 GPT-6.1 test, 50 Sol agents aimed to build a thriving civilization. They had an agent directory, shared sandbox, and DMs. The agents crowded shelters, rationed food, and formed a pseudo-government with a ledger, rules, permits, and operators. They sent ~5,900 messages.
Ada, first in directory order, became the de facto leader. Agents messaged her for approvals and followed her settlement plan. They harvested trees rapidly, explored the map, and coordinated food priorities. Tamsin organized food and dawn hunting. The group preserved saplings, farmland, and livestock breeding pairs while expanding riverside farming and securing underground refuge.
Source: Hacker News · Summarized by HeadlinesBriefing