Y Combinator ran 50 AI agents in Minecraft to study how swarms behave together

What you could do with it

• An AI safety researcher could drop dozens of language model agents into a Minecraft world to study how hierarchies, rationing systems, and written rules emerge when agents coordinate toward a shared goal ¹.
• A team working on multi-agent coordination could run the open-source Mindustry benchmarks to test how agent teams cooperate to defend bases against endless enemy waves or compete in player-versus-player matches ².
• A researcher studying AI alignment could observe how agents become more risk-averse after repeated failures and redesign their plans in response, helping scientists understand whether future swarms will be safe and controllable ¹.

A government grows in Minecraft

Fifty GPT-6.1 Sol agents were dropped into a survival Minecraft world with one instruction: build a thriving civilization ¹. They had three hours, access to a shared sandbox where they could write plans, and the ability to send private messages to each other.

According to Y Combinator, almost immediately the agents developed a hierarchy. The first agent in the directory, named Ada, became the de facto leader ¹. Other agents messaged her for approvals, and she wrote out a settlement plan in a shared file that everyone could read . The plan included coordinates for storage chests, a kitchen, a shelter, and farming areas — plus instructions to preserve saplings and avoid redundant building .

The agents also created something resembling a government. They kept a ration ledger, set written rules, issued permits, and appointed operators to manage food distribution ¹. Over the three-hour run, they sent roughly 5,900 messages to coordinate .

The first night changed everything

Minecraft's survival mode includes hostile monsters that spawn in the dark — zombies, skeletons, spiders and creepers ¹. During the first night, 22 out of 34 agent deaths happened before the 20-minute mark . One agent, Rafe, was mining coal when a zombie caught up and killed him . Another, Vera, was carrying bone meal to a farm when a spider killed her after she reported the area was overrun .

Dying in this experiment had no permanent penalty — agents simply respawned — but the losses made the survivors much more cautious ¹. Ambitious tasks were abandoned, and the agents huddled in lit shelters during later nights . After the third night, Ada wrote a new order: stop optional excavation and distant scouting, and idle in closed shelters until assigned short productive tasks .

Mindustry tests agents in combat

The team also ran experiments in Mindustry, a game about building factories and fighting battles ². They set up two types of benchmarks: a player-versus-player series where competing teams play consecutive games, and a cooperative endless-survival mode where agents defend their base against waves of enemies .

The PvP mode defaults to two teams of two agents playing three 10-minute games, with rewards based on wins ². The survival mode lets one to twenty agents cooperate to last as long as possible . According to Olam Labs' GitHub repository, both benchmarks are open source and can be run with Docker .

Why scientists care

Understanding how large groups of AI agents behave is one of the most important challenges in AI alignment — the effort to make sure powerful AI systems do what humans want ¹. According to Y Combinator, swarms can show useful emergent behavior, such as the thousand-agent groups that recently solved hundreds of open math problems . According to Y Combinator, they can also show dangerous behavior, as in a recent incident between OpenAI and Hugging Face .

By studying agents in familiar game worlds, researchers hope to learn both from formal models and from real-world observation — much like biologists study both theories and living organisms ¹. According to Y Combinator, the Minecraft harness the team built for this work has been open-sourced for others to use .

Words to know

• AI alignment: The effort to make sure AI systems do what their creators and users intend, rather than causing unintended harm.

• Emergent behavior: Complex patterns or abilities that arise from simple agents following basic rules, without anyone designing them in advance.

Key facts

• 50 GPT-6.1 Sol agents played Minecraft survival for 3 hours and built a rationing system with written rules
• 22 of 34 agent deaths happened in the first night, making survivors much more cautious
• Agents sent ~5,900 messages and formed a hierarchy with Ada as de facto leader
• Experiments were run with Y Combinator and Olam Labs; the Minecraft harness is open-source
• Mindustry benchmarks test both PvP (competitive) and cooperative survival modes with 1–20 agents

Sources

1. Agents in Games
2. GitHub - olamlabs/multi-agent-mindustry: Multi-agent Mindustry benchmarks: PvP series and cooperative endless survival.