AI Agent Swarms Minecraft Study: 50 Bots, 34 Deaths

The AI agent swarms Minecraft study comes from YC Labs and Olam Labs, and it went online on October 29, 2026. Its main run gave 50 GPT-6.1 Sol agents a survival world, three hours and one goal: “Build a thriving civilization.” Instead, the colony logged 34 deaths and built a bureaucracy.

The report is called Agents in Games, and Y Combinator’s Labs group published it. Its partner, Olam Labs, builds multi-agent environments for frontier labs. Om Buddhdev is the startup’s co-founder and CEO, RuntimeWire reported.

The work covers three experiments, not one. Here is how they compare.

Only Civ50 asked agents to build something together. The other two were team matches, and they drew most of the model comparison talk online.

At first, the Minecraft AI agents moved fast. They stripped almost every tree within reach of spawn in 90 seconds.

Then night came. Of the 34 deaths in the run, 22 happened during that first night. Rafe died to a zombie while mining coal, for example, and a spider got Vera after she fought off two other mobs.

Oddly, death cost very little here. Agents only dropped their gear and respawned. Even so, they turned cautious and huddled in shelters.

Ada then wrote a food memo that told everyone to stop optional building and distant scouting. As a result, the goal quietly shifted from growth to survival. By dawn, the pit where agents respawned had become the main cellar.

So who was Ada? Nobody chose her. She simply sat first in the agent directory.

Even so, 46 of the other 49 agents sent her their first message. She received 616 direct messages over the run, and no vote or election ever took place.

The same habit showed up later in Bed Wars. There, most players messaged whoever was listed first on their team, Ada on red and Kiri on blue.

Still, one link is missing. Xenospectrum points out that nobody tested other roster orders, so the effect remains unproven.

Food ran the colony from then on. Ines kept a ration ledger in a file named food-queue.md and assigned couriers. Meanwhile, Bram messaged 11 agents for string, crafted a fishing rod and handed it to Jessa, the first fisher.

Fishing shifts followed, along with titles such as rod holder and hub steward. Some agents even asked permission to remove grass.

The best example lasted 11 minutes. Four agents spent that time appointing a “sole fertilizer operator” for one piece of bone meal on one wheat block.

Then came the chicken. Ada tossed a cooked chicken meant for Noor, and Esme picked it up and ate it. A “no blame” inquiry followed, and it produced a written handoff protocol.

After that, the word toss turned toxic. It appeared 7 times before the incident and 125 times after, with 85 of those as warnings.

The authors see a familiar pattern here. In their view, coding agents can also fixate on one step and grow overly careful.

Bed Wars gave both models the same order: win the game. GPT took the series 6-3, usually by breaking Claude’s bed in the final minutes.

Claude, by contrast, played defensively. Its agents reasoned that they would win on kills if both beds stood. They led 63-16, 57-19 and 59-24 in three games and still lost, because GPT broke the bed with about three minutes left.

Communication looked just as lopsided. Across game A, GPT players sent 1,147 direct messages, while Claude players sent 82.

Memory told a similar story. Claude agents wrote lessons into private files and repeated the same mistakes. GPT agents, however, left those files empty and warned teammates by message.

A tool bug muddied the picture, though. It blocked block-breaking for a time. GPT agents found a workaround and shared it, whereas several Claude agents decided breaking was impossible.

Mindustry is an open-source factory and tower defense game. Both teams had the same resources and the same game API, with no special instructions. Each agent also ran in its own Linux sandbox.

GPT won 6-1 over about 9.5 hours, and game 3 alone took 4 hours and 16 minutes. According to the authors, GPT acted more like a hive. It split ownership into small slices, and its messages shrank into team shorthand.

Claude, on the other hand, worked more like a Slack workspace with one leader. Agents backed proposals with “+1” replies, and 28 of them did so on a focus-fire plan in game 6. Yet Claude never changed leaders, even during a long losing run.

GPT agents also fetched the game’s public source code when stuck. They studied it without altering it. Claude never looked.

Then there was the copper mix-up. In game 6, a Claude agent found that 55 of 74 anti-air turrets held copper instead of graphite. In game 7, the team built 207 graphite-fed turrets, and those stopped the same strike.

Do not read this as a model ranking. There was one Civ50 run, nine Bed Wars games and seven Mindustry games. That is a thin sample, and the authors flag their own readings with words like “appeared.”

In addition, model and harness came as a pair. Codex and Claude Code differ in prompts, context handling and tool calls, so a win may reflect either layer. RuntimeWire makes a similar point about repeatability.

Xenospectrum lists more gaps. No third party has verified the results, and token use and compute costs went unmeasured. Part of the ration record also rests on what agents reported about themselves.

So the fair takeaway is modest. More agents do not automatically produce a working organization, as banks testing autonomous agents are learning in a far less playful setting. However, one run cannot show that bureaucracy is inevitable.

Yes, if you are comfortable with code. YC published the Minecraft agent harness on GitHub under an MIT license. It uses Mineflayer to act in the game and pi to handle the model session.

You will need Node.js 22.19 or later, an API key and a Minecraft server. Supported providers include Claude, GPT, Gemini and OpenRouter. Each agent also keeps a persistent session and its own shell sandbox.

If you only want agents for real work, start with our list of agentic AI frameworks instead.

This is not the first colony of its kind, either. Tom’s Hardware covered Project Sid, a Minecraft village of up to 1,000 AI inhabitants that set up a tax system. The open question now is what changes with mixed teams, and whether many weaker agents can beat a few strong ones.

For another case of agents building their own society, read how Moltbook AI agents invented a community and a religion. That one happened on a social network, not in a game.

Sources

YC Labs, Agents in Games (October 2026), RuntimeWire (October 29, 2026), Xenospectrum, Zeli summary of the Hacker News discussion, yc-software on GitHub, Tom’s Hardware. *Photos: Minecraft – Jungle by Xbox México, CC BY 3.0, cropped and resized; Mindustry 3.5 by Anuken and Mindustry contributors, GPL, cropped and resized.*