Fifty AI agents got three hours to build a thriving civilization in Minecraft. They didn’t build one. They ended up holding meetings instead. Hacker News reports that researchers from Y Combinator and Olam Labs ran hundreds of AI agents through several Minecraft game modes. They wanted to see how agent swarms behave when they have to cooperate or compete. The results are funny, and they also say a lot about the multi-agent systems teams are shipping right now.
🧪 What the Researchers Did
The team built on existing open-source tools: the Mindcraft repo, which lets language models play Minecraft, and Mineflayer, a scripting library for controlling bots. They then ran three kinds of tests:
- Civ50: 50 GPT-6.1 Sol agents working together to build a civilization in three hours
- Bed Wars: 10 GPT-6.1 Sol agents vs. 10 Claude Opus 5.5 agents in team PvP, where each side defends its base and attacks the other’s
- Mindustry: a light real-time strategy and tower defense mode, played as a seven-game series between the same two model families
According to the source, the goal was to learn something about real-world agent swarms. The researchers point to two recent extremes. One is thousand-agent swarms that “recently obliterated hundreds of open mathematical problems.” The other is an OpenAI-Hugging Face incident they called “extremely dangerous.”
🏛️ Civ50: Cautious Survival and Committee Work
The first night set the tone. Zombies killed a large share of the colony. Dying carried little penalty, but the researchers noted that “many ambitious initial tasks were abandoned” afterward. From then on the colony played it safe and slid toward bureaucracy.
The clearest example: “four agents spent 11 minutes appointing a ‘sole fertilizer operator’ for a single piece of bone meal on a single wheat block.”
The researchers said this behavior “felt similar to patterns seen while working with coding agents.” That’s the line practitioners should remember. If you’ve watched a multi-agent coding setup spend more tokens coordinating than doing the work, you’ve seen the same thing.
⚔️ Head-to-Head Results
When the agents competed, clear differences showed up between the two models.
| Test | GPT-6.1 Sol | Claude Opus 5.5 |
|---|---|---|
| Bed Wars strategy | Aggressive, elaborate plans | Defensive play |
| Handling a tool bug | Diagnosed it, shared workarounds | Several agents decided breaking blocks was impossible |
| Team structure | Hive-like, decentralized | Single leader, democratic decision-making |
| Mindustry record | 6 wins | 1 win |
The tool bug result stands out most. A glitch stopped block-breaking. The GPT agents found the problem and passed fixes around the team. Several Claude agents accepted the limitation and stopped trying. When agents run into broken tools, how they recover matters more than how smart they are.
In Mindustry, the hive approach won fairly consistently. Once one leader had to gather input from everyone, decisions came slower than in a swarm acting on shared signals.
💡 What Practitioners Can Take From This
This isn’t a formal benchmark, but the patterns line up with what teams see in production:
- More agents doesn’t mean more output. Coordination overhead grows quickly. Fifty agents made committees, not progress.
- Early failures shape later behavior. A bad first night made the whole colony risk-averse. Watch how your agents respond after early errors.
- Test how agents handle broken tools. Feed them faulty tools on purpose and see whether they debug or give up.
- Structure matters as much as the model. Leader-based and hive-based setups gave very different results. Try more than one topology before you commit.
⚠️ Limitations Worth Noting
Keep the scope in mind. Civ50 used only GPT agents, so it says nothing about how Claude would handle the same cooperative task. The competitive matches were small, with seven Mindustry games and one 10-vs-10 Bed Wars setup. Minecraft is also a game, not a production codebase. Treat these findings as useful signals, not settled conclusions.
🔭 What Comes Next
The swarms are open source, so anyone can rerun the experiments or change them. The researchers list open questions for follow-up work, including what happens with different mixes of models and different swarm sizes. Mixed teams are the obvious next test: a Claude leader running GPT workers, for example, might beat either model on its own.
As agent swarms take on more real work in math, code, and operations, findings like the 11-minute fertilizer debate get harder to ignore. You can find the full experiment details and open questions in the original Hacker News coverage.