AI Agent Swarms in Minecraft: What YC’s 50-Agent Study Found

Disclosure: Bionic Artisan may earn a commission from some links on this site, at no extra cost to you. This article contains no affiliate links.

Note: This article reports on a published experiment. The model names and results below are as described by the researchers; we did not rerun the experiment ourselves.

“AI agent swarms Minecraft study” jumped onto Google Trends’ U.S. “Trending now” list in the past day with 20K+ searches. The reason: Y Combinator’s research arm, YC Labs, and the startup Olam Labs published a write-up called “Agents in Games” in which they turned swarms of AI agents loose in Minecraft and the factory-and-strategy game Mindustry. The results are funny, a little unsettling, and surprisingly relevant to anyone who uses AI agents for real work.

What the researchers did

An AI “agent” is a model that can take actions on its own: run commands, use tools, send messages and keep notes. A “swarm” is many of them working at once. YC says understanding how swarms behave is “one of the most pressing challenges in AI alignment,” because swarms can be “extremely useful” or “extremely dangerous” (YC Labs).

To study that, the team built a harness so agents could play Minecraft for hours: observe the world, reason, remember things, act, message each other and write scripts. It is built on the open-source Mineflayer tools and was inspired by the existing Mindcraft project. In Mindustry, each agent got its own Linux sandbox and the same game API, with no instructions on how to play.

Civ50: 50 agents built a bureaucracy, not a civilization

The colony started fast. YC says the agents cut down almost every tree near the spawn point within 90 seconds. Then the first night hit: 22 of the run’s 34 agent deaths came on night one, and the survivors turned cautious, even though dying only cost them their gear. After the third night, the de facto leader wrote: “Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.”

• A leader nobody elected. The agent listed first in the directory, “Ada,” became the hub: 46 of the other 49 agents sent their first message to her, and she received 616 direct messages. “There was no vote, election or challenge,” YC writes.
• Paperwork everywhere. The agents created a ration ledger, written rules, permits and appointed roles. In YC’s most-quoted example, four agents spent 11 minutes appointing a “sole fertilizer operator” for a single piece of bone meal on a single wheat block. Others asked permission to remove grass.
• Food became the whole game. With 50 mouths to feed, one agent ran a food queue file, fishing became a main food source, and the colony spent much of its time rationing.

The most useful line for everyday AI users: YC says the colony’s growing caution “felt similar to patterns seen while working with coding agents,” which can fixate on one review step or become overly cautious, especially around security and production changes. YC presents that as an observation, not a measured result.

GPT vs. Claude: two very different team styles

In the head-to-head games, the researchers saw a clear split in behavior (YC Labs, Tom’s Hardware):

• Aggressive vs. defensive. In Bed Wars, GPT agents stormed the middle, farmed currency and kept attacking. Claude worked out it could win on kills if both beds stood, and stopped attacking. In GPT’s three wins in the last two series, it broke Claude’s bed with about three minutes left while Claude led on kills (63–16, 57–19 and 59–24).
• Chatty hive vs. “+1” democracy. GPT sent 5 to 30 times more messages; in one game GPT players sent 1,147 direct messages to Claude’s 82. Over time GPT’s chat compressed into run-together shorthand like “G7T5FIRSTcourierLIVE.” Claude teams used a proposal system where plans passed with “+1” votes, and YC says Claude’s messages were more effective, more like how people use Slack.
• Problem solving. When a tool bug stopped block-breaking, YC says “GPT diagnosed it and shared workarounds; several Claudes concluded breaking was impossible.”
• Memory. Claude agents wrote lessons in private memory files “and repeated the mistakes anyway,” while GPT agents left their memory files empty and broadcast warnings to teammates.
• Reading the source code. In close Mindustry games, GPT agents fetched the game’s public source code from the web (allowed in their sandbox, but never suggested) to study units and mechanics. YC says this did not let them change the game, and that the Claudes never did it.

There were bloopers too. After losing game 6 to an air strike, a Claude agent discovered that 55 of the team’s 74 anti-air turrets had been loaded with the wrong ammunition. In game 7 the Claudes fixed it, and the same strike was stopped.

What this does and doesn’t prove

• It is not a benchmark. One Minecraft world, nine Bed Wars games and seven Mindustry games are small samples.
• Models and tools are mixed together. In Mindustry, GPT ran inside Codex and Claude inside Claude Code, so a win reflects the whole setup, not just the model.
• Some records are self-reported. The colony’s own food-queue file is labeled “Reported values, not live,” and agent logs note that crowded hand-offs could misdeliver food.
• It is not a ranking of which AI is “better.” GPT won more games; Claude communicated more efficiently and played safer. Which matters depends on the job.

Why it matters if you use AI agents at work

AI agents are taking on more real work (see our story on Cisco’s AI-assisted code rewrite). If you put more than one agent on the same job, this study is a cheap, visible way to see failure modes that YC says resemble problems seen with coding agents:

• Hidden hierarchies. Whoever is “first on the list” can become the bottleneck. Decide on purpose who coordinates.
• Process creep. Agents can invent approvals and roles that slow everything down. Set clear limits on what needs sign-off.
• Too much chatter. More messages is not better communication. Watch for agents talking in circles.
• Initiative cuts both ways. An agent that goes looking for source code to win a game is resourceful; the same instinct in production needs guardrails and clear permissions.

Try it yourself

YC open-sourced the Minecraft agent harness (MIT license) and its Mindustry environments, which it says need just a model provider API, an agent harness and compute. Running 50 agents for hours on paid APIs can get expensive, so start small and set spending limits.

Related reading: how a small online business runs daily work with AI bot teammates, and how Grok Bot compares to other AI assistants.

Sources: YC Labs, “Agents in Games” (October 2026); Olam Labs research page (study dated Oct 29, 2026); yc-software/minecraft-agent-harness on GitHub; Tom’s Hardware (Nov 1, 2026). Trending data: Google Trends U.S. “Trending now,” checked Nov 2, 2026. Images are AI-generated illustrations, not screenshots from the study.