Y Combinator's AI agent swarms in Minecraft and Mindustry, explained: what 50 agents and two rival teams show about running agents in groups

Y Combinator and Olam Labs put 50 AI agents into one Minecraft world, then pitted GPT-6.1 Sol teams against Claude Opus 5.5 teams in Bed Wars and Mindustry. We pull the numbers from all three runs into one place and turn each failure into a check for teams running several coding agents at once.

This explains reporting by

Y Combinator Labs, "We ran AI agent swarms in Minecraft and Mindustry" (October 2026). Read the original first:

https://labs.ycombinator.com/projects/games/index.html

In one minute #

• Civ50: 50 GPT-6.1 Sol agents spent 3 hours in one Minecraft survival world, told to build a thriving civilization. They sent about 5,900 messages and died 34 times, 22 of those in the first night.
• The first agent in the directory, Ada, became leader with no vote: 46 of the other 49 sent her their first message, and she received 616 direct messages.
• The colony drifted from building to survival and process. Four agents spent 11 minutes appointing a sole fertilizer operator for one piece of bone meal.
• Head to head, GPT-6.1 Sol beat Claude Opus 5.5 6 to 3 over nine Bed Wars games and 6 to 1 over seven Mindustry games. Together that is 12 to 4.
• GPT teams messaged far more, read Mindustry's source code and shared a fix for a tool bug. Claude agents kept one leader, wrote lessons to memory files and repeated the same mistakes.
• The Minecraft harness is open source, so the setup can be rerun.

What YC ran #

Y Combinator ran the experiments with Olam Labs, which builds multi-agent environments for AI labs. The page frames group behavior as an alignment problem: what agents do together that no single agent was told to do.

There were three setups. Civ50 put 50 GPT-6.1 Sol agents in one Minecraft survival world for 3 hours, with an agent directory, a shared sandbox and direct messages. Bed Wars ran 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents for nine 45-minute games, each team told to win. Mindustry, a factory-building and tower-defense game, pitted GPT-6.1 Sol in Codex against Claude Opus 5.5 in Claude Code, with the same game API and no strategy hints.

The Minecraft harness runs its core loop on Pi, with Mineflayer tools for acting in the world, and is published at github.com/yc-software/minecraft-agent-harness.

Civ50: a leader chosen by list order #

Nobody held an election. Ada was first on the roster, and 46 of the other 49 agents sent their first message to her. She received 616 direct messages over the run.

Agents died 34 times, 22 of them in the first night. After that the colony grew cautious, took on fewer risky tasks and turned to food and safety. Fishing became the main food source, with about a dozen agents taking shifts and a cook-and-deliver system around it.

They invented job titles such as fishing shift and hub steward. Four agents spent 11 minutes appointing a sole fertilizer operator for one piece of bone meal. When a cooked chicken tossed to one agent was eaten by another, they adopted a rule: named chest withdrawals, no tossing.

The head-to-head record #

Game        Teams                                   Result (GPT-6.1 Sol vs Claude Opus 5.5)
Bed Wars    10 vs 10, nine 45-minute games          6 - 3
Mindustry   Codex vs Claude Code, 7 games, ~9.5 h   6 - 1
Total       16 games                                12 - 4

• Bed Wars: Claude played defensively for kills and led on kills in the last series (63 to 16, 57 to 19, 59 to 24), but GPT broke Claude's bed about three minutes before the end in its three wins there.
• Mindustry: game 3 alone ran 4 hours 16 minutes, and most games except the last two were close.
• This is two game types, one harness per game and one set of instructions, and the Mindustry teams ran in different agent tools. It says more about these setups than about the models in general.

How the two teams talked #

In Bed Wars, GPT agents sent 5 to 30 times as many messages as Claude agents. In one game GPT sent 1,147 direct messages to Claude's 82. GPT's most-messaged player received 198 to 368 a game; Claude's received 11 to 48.

Claude agents said we in about 30% of messages, GPT agents in about 5%. In Mindustry, GPT's messages became compressed, with words run together, while Claude's read like a human team chat.

Claude used one leader and a +1 vote on proposals, and never changed leaders, even while losing. GPT split ownership into small slices, and its agents took the initiative when a game got close.

Where the teams learned, and where they did not #

A tool bug blocked breaking blocks in Bed Wars. GPT agents worked out that equipping the pickaxe explicitly got around it and shared the fix. Several Claude agents concluded breaking was impossible.

Claude agents wrote lessons into memory files and still repeated mistakes such as falling off edges. GPT agents left their memory files empty and warned each other by direct message instead.

In Mindustry, GPT agents fetched the game's public source code to study units and mechanics; Claude agents never did. In game 6 Claude's team loaded 55 of 74 anti-air turrets with copper instead of graphite. In game 7 it built a ring of 207 graphite-fed turrets that stopped the same air strike.

What this means if you run several coding agents #

The authors compare the caution they saw to coding agents that fixate on one step or hold back too much on security and production changes. The checks below are our reading of the results, not YC's:

• A leader by list order: if every agent reports to whoever is listed first, that agent becomes the bottleneck. Pick the coordinator on purpose and watch how many messages it receives.
• Bureaucracy: give agents a time or token budget per decision, or they may spend it on process.
• Roles that never change: let a stuck team replace its lead, or have a supervisor do it.
• Memory that does not change behavior: a lessons file only helps if the next step reads it. Test that it does.
• Reading the source: the team that read the game's code did better. Give coding agents the docs and source of the tools they use.

Who is affected #

What to do #

• Read the YC page's per-game notes before quoting the score.
• Clone the open-source harness to rerun Civ50 on other models.
• In your own multi-agent setup, choose the coordinator on purpose and log how many messages it gets.
• Check that your agents' memory or lessons files are actually read before each task.

What is still unknown #

• Each result we saw is from one run. The page does not report repeat runs of Civ50.
• Prompts, reasoning settings and agent tools differ between teams in places (Codex against Claude Code in Mindustry), so model and harness effects are mixed.
• Whether these patterns show up in coding work is the authors' comparison, not a measurement.

Sources #

• Y Combinator Labs, "We ran AI agent swarms in Minecraft and Mindustry" (October 2026) — the original report
• Hacker News discussion of the YC page

AI News Report· every headline, every morning.