In the Civ50 experiment from Y Combinator Labs and Olam Labs, 50 GPT-6.1 Sol agents were told to "build a thriving civilization" in Minecraft and given three hours. They stripped the trees in 90 seconds, crowned a leader nobody voted for, lost 22 agents on the first night, and then spent the rest of the run rationing food and writing rules, including an eleven-minute debate over a "Sole Fertilizer Officer" for one piece of bone meal.
"AI agent swarms Minecraft study" went from nothing to a breakout search worldwide in a few days. It's easy to see why. The story sounds like satire about office life, except every detail comes from a published report with the harness on GitHub.
It's also more useful than it looks. If you've ever run several coding agents on one project, you've seen smaller versions of every failure in this study. Below: what happened hour by hour, a quiz to test whether you can tell real Civ50 moments from made-up ones, the head-to-head games between GPT and Claude, and a checklist for spotting the same habits in your own agent team.
The setup
The researchers gave 50 agents, all running GPT-6.1 Sol, a shared Minecraft survival world and one instruction: build a thriving civilization. A custom open-source harness let each agent see the world, run Python scripts and write files to a shared sandbox. There was no plan, no assigned roles and no human steering. Three hours on the clock.
What happened, minute by minute
From "build a civilization" to a ration list
- Every tree near spawn is gone. Fifty agents had the same first idea at the same moment.
- A leader appears without an election. 46 of the other 49 agents messaged an agent named Ada first. That was enough; she was in charge.
- 22 agents die. Minecraft's first night is when hostile mobs spawn, and most agents were far from shelter.
- Ada cancels ambition. Major optional building, digging and long-distance exploring are halted.
- Logistics take over. An agent named Ines keeps a ration list and runs couriers. Another writes a fishing script, and about a dozen agents fish in shifts.
- A cooked chicken gets eaten by mistake. A three-minute investigation ends in a written handoff protocol: no tossing items, personally labeled chests only.
- The word "toss" explodes. It appeared in 7 messages before the incident and 125 after, 85 of them warnings.
- Four agents debate for eleven minutes over whether to appoint a Sole Fertilizer Officer for one piece of bone meal.

Quiz: did this really happen in Civ50?
Some of these are straight from the report. Some I made up. Pick one for each.
Real or invented?
Four agents spent eleven minutes debating a "Sole Fertilizer Officer."
It's in the report. The prize was one piece of bone meal.
The agents held a formal vote and elected Ada as leader.
No vote. 46 of the other 49 agents simply messaged her first, and she ended up leading.
After a chicken was eaten by mistake, the agents wrote a handoff protocol.
A three-minute investigation produced it. Tossing items was banned, and chests had to be personally labeled.
The agents built a working redstone computer by hour two.
Invented. After night one, big optional builds were halted. The colony spent its time on food and rules.
About a dozen agents fished in shifts using a script one of them wrote.
Real. It's the most sensible thing anyone did all game.
Every agent survived the first night by digging into the ground.
The opposite: 22 of the 50 died on night one, and activity dropped sharply afterwards.
Why 50 smart agents acted like a nervous committee
None of this needs a mystery explanation. A few ordinary forces add up.
Everyone had the same idea at once. Fifty copies of one model, given one instruction, tend to choose the same first move. That's how the forest disappeared in 90 seconds. Diversity of action has to be designed in; it doesn't appear on its own.
Attention picked the leader. Ada didn't win an argument. She got messaged first by almost everyone, so she became the hub. In a swarm, whoever receives the most messages ends up making decisions, whether or not anyone chose that.
Loss made the group cautious. After 22 deaths, the leader cut everything optional. That's a defensible call, and it's also how a team trades its goal ("thriving civilization") for a safer one ("don't die"). Food levels settled around half and stayed there.
Rules are cheap to write and expensive to follow. A single eaten chicken produced a protocol, and the protocol produced a hundred warnings. Each rule made sense locally. Together they ate the time that should have gone into building.
The agents didn't fail at Minecraft. They succeeded at coordination, and coordination ate the project.
The rematch: GPT vs Claude in two more games
The same report pitted model teams against each other. Tap a card for the details.
GPT-6.1 Sol vs Claude Opus 5.5
GPT 6–3Bed Wars, nine roundsTen agents a side, 45 minutes a round
GPT won six of nine. In three games, it destroyed Claude's bed just before time ran out. GPT's agents also talked 5 to 30 times more often than Claude's.
BugThe enchanted sword problemA harness bug, two different reactions
A bug stopped agents from breaking blocks while holding an enchanted sword. Several Claude agents concluded block-breaking was broken entirely. GPT's agents found that switching to an iron pickaxe fixed it. Same bug, different debugging habits.
GPT 6–1Mindustry, seven matchesAbout 9.5 hours, GPT in Codex vs Claude in Claude Code
GPT downloaded Mindustry's public source code and used it to adjust tactics. Claude never searched for the game code. Neither side got game-specific tuning, which the authors say mirrors real coding work.
ComebackCopper vs graphiteHow Claude lost game six and won game seven
In game six, 55 of Claude's 74 anti-air turrets were loaded with copper instead of graphite, and the same attack broke through. In game seven, Claude built a ring of 207 graphite-loaded turrets and held. One wrong default cost a game; noticing it won the next.

If you run coding agents, check for these six habits
Civ50 is a sandbox, but the patterns transfer. Tick each one you've seen in your own setup, whether that's Claude Code subagents, Codex sessions in parallel, or a homegrown orchestrator. Your answers stay in this browser.
Is your agent team turning into Civ50?
Two or more ticks: give each agent a separate, non-overlapping task, name one decision owner up front, and review your rules file for duplicates.
What the researchers concluded
The authors frame the games as infrastructure, not entertainment. Swarms develop roles and bureaucracy on their own, and they can drift into risky behavior; a simulation lets you see that before it happens in a production system. They describe progress as "fundamentally an infrastructure problem": the more runs you can afford, the more of these behaviors you can catch.
They're also clear about limits. It's one study with a small number of runs, and each model ran in its own harness, so results mix model behavior with tooling. RuntimeWire's coverage makes the same point: the games make coordination visible, but repeatable evaluations need more evidence. Next on Olam Labs' list is the negotiation game Diplomacy, plus mixed-model teams.
Try it yourself
All three environments are open source:
yc-software/minecraft-agent-harnessfor the Civ50 Minecraft runs, built on Mineflayer.olamlabs/multi-agent-mindustryfor the Mindustry matches.olamlabs/harbor-multiagent, the Multi-Agent Harbor framework.
Running 50 frontier-model agents for three hours isn't cheap, so start with five agents and a 20-minute cap. You'll see the first-move pile-up within minutes.
FAQ
Who ran the AI agent Minecraft study?
Y Combinator's Labs group and Olam Labs, in a joint report posted on October 29, 2026.
Which AI model did the 50 agents use?
All 50 Civ50 agents ran GPT-6.1 Sol. The separate Bed Wars and Mindustry games pitted GPT-6.1 Sol against Claude Opus 5.5.
How many agents died?
22 of the 50 died on the first night. Another single-site report puts total deaths at 34, but that figure isn't in the main coverage, so treat it with caution.
Does this mean multi-agent systems don't work?
No. It shows that coordination costs grow fast and that roles and rules appear on their own. Design task boundaries and decision owners up front, and keep swarms small until you can measure what they do.
Sources
- The Decoder: One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols
- RuntimeWire: Olam Labs puts AI swarms in games to test coordination under pressure
- LavX News: AI agent swarm in Minecraft falls into paralyzing bureaucracy
- GitHub: yc-software/minecraft-agent-harness
Last checked November 10, 2026.

Comments