AI Agent Swarms Run Wild in Minecraft and Mindustry

Researchers involved in a collaborative project described AI agent swarms that played games. The team worked with Olam Labs and used GPT-6.1 Terra and GPT-6.1 Sol agents in Minecraft and Mindustry, according to a report on Hacker News Labs, and the work matters for AI alignment.

Video file

Video file

What the researchers tried to learn

The report framed multi-agent dynamics as a pressing challenge. It said swarms working toward shared goals can develop roles, hierarchies, and coordination patterns. These may be useful, as with large swarms that tackled open mathematical problems, or dangerous, as in an OpenAI-Hugging Face incident mentioned in the source. The researchers wanted empirical data. They built a harness for long-running agents that could observe, reason, remember, act, message, and execute scripts. According to the source, the harness was built on Mineflayer and Pi, and the team said it open-sourced the work.

How the Minecraft swarm was organized

The source described a new Minecraft colony with 25 GPT-6.1 Terra agents, with a follow-up on Twitch. It also described Civ50, a run with 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours, with one goal: 'Build a thriving civilization.' The agents had an agent directory, a shared sandbox, and DMs. They sent about 5,900 messages. They formed a pseudo-government with a ration ledger, written rules, permits, and appointed operators.

Initially, the agents were active. They spread out and explored. They were fixated on harvesting trees. They harvested almost all trees near spawn within 90 seconds. A hierarchy formed based on directory order. Ada, first, became a de facto leader. Agents messaged her for approvals, and she wrote a plan in the shared sandbox.

The first night brought deaths. The report said 22 of the run's deaths occurred in the first night before minute 20. Many agents became cautious. After the third night, Ada wrote for agents to stop optional bulk excavation and building, and to idle in lit closed shelters until assigned short productive tasks.

What happened under food and survival pressure

The run shifted from a potentially thriving civilization to one focused on food and survival. The agents formed a food rationing system routed through Ines. She handled requests and tracked food in a shared file. The report included messages and ration calculations. Agents had targets, hunger values, and food restoration values. They rationed cooked food, bread, apples. They stopped distant scouting and bulk building.

The source said 50 agents meant many mouths to feed. The agents became risk-averse, huddled in shelters, and attempted fewer risky tasks. This showed coordination can emerge, but also how resource pressure shapes behavior.

What the videos and open questions suggest

The source page included two embedded videos: team chat and zenith strike. The team chat clip matters for the messaging and ration coordination. The zenith strike clip matters for the more action-oriented swarm behavior in Mindustry. The lab listed open questions, funny bloopers, and a run-your-own-swarm guide.

How Grafy relates

Grafy's Game Studio at /game is the most genuine connection. It turns a brief into a playable game, which could let creators prototype a swarm scenario. Grafy Chat at /chat can bring multiple AI models into one workspace for role-based prompting. Grafy Graph at /graph can branch prompts, outputs, and model history. The source does not mean Grafy has an agent swarm product. But the underlying questions about coordination and roles are relevant.

Frequently asked questions

What was the main goal of the Minecraft swarm run?

The researchers wanted to study how many AI agents coordinate toward a shared goal. According to the source, they gave agents survival tasks like gathering food, building shelter, and exploring. The Civ50 run asked for a thriving civilization. The team wanted to see roles, hierarchies, and coordination emerge.

Did the agents show emergent coordination?

The report described an hierarchy and rationing system. Ada became a de facto leader, and agents sent thousands of messages. After early deaths, they became cautious and stopped risky tasks. The source said swarms can be useful or dangerous, and the experiment added empirical observations.

How can developers explore agent swarms?

The lab said it open-sourced the agent harness, according to the source. Developers could inspect how agents observed, remembered, and acted. In Grafy, a reader might use Game Studio to prototype a swarm-like game or Chat to assemble model roles. The source does not show a direct swarm app in Grafy.

---

Source: Hacker News — front page — We ran AI agent swarms in Minecraft and Mindustry