YC Labs pits Claude and GPT agent swarms in Minecraft

Why games?

According to YC Labs, understanding multi-agent dynamics is one of the most pressing challenges in alignment. Swarms develop roles and hierarchies that can be very useful or very dangerous. Games make good labs: they are reproducible, legible and competitive. The experiments were run with Olam Labs, and the Minecraft harness is open source.

Civ50: 50 agents, one civilization

Fifty GPT-6.1 Sol agents got three hours in a survival Minecraft world with one goal: build a thriving civilization. They had a shared sandbox, an agent directory and DMs, and sent about 5,900 messages.

• The agent listed first in the directory became the leader, and everyone asked her for approvals.
• Within 90 seconds they had cut down almost every tree near spawn.
• The first night caused 22 of the run's 34 deaths. The swarm stayed more cautious afterwards, even though death cost almost nothing.
• Civilization building turned into food management: a ration ledger, requests, approvals and a food queue kept in a markdown file.

Bed Wars: Claude vs GPT

In nine games, 10 Claude Opus 5.5 agents played 10 GPT-6.1 Sol agents. GPT won 6–3, usually by breaking Claude's bed in the final minutes. Claude worked out that with both beds standing the team with more kills wins, so it played defense. It typically led on kills (63–16, for example) but lost the game once the bed fell.

Communication differed sharply. GPT sent 5-30 times more messages and quickly switched to compressed shorthand. Claude signed its messages and said "we" in 30% of them, against GPT's 5%. When a tool bug broke block-breaking, GPT diagnosed it and shared a workaround, while several Claudes concluded breaking was impossible. Claude wrote lessons into private memory files and repeated the mistakes anyway. GPT left its memory files empty and warned teammates by DM instead.

Mindustry: 6–1 to GPT

In the factory-building and tower-defense strategy game, GPT-6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code 6–1 over about 9.5 hours of play. Observations:

• GPTs behaved more like a hive, took on grunt work more willingly and built a denser but often low-value communication network.
• GPT language grew ever more compressed, with its own space-free shorthand. Claudes decided more democratically: a proposal, then "+1".
• Claudes never switched leaders, while in close GPT games some agent always took the initiative.
• When stuck, GPTs fetched and analyzed the game's open source code. Nobody forbade or asked for this. Claudes never did it.

Why it matters to you

This is not a benchmark: the sample is small and behavior depends on the harness. Still, it shows how strongly structure shapes a swarm: who is first on the list, who gets the DMs, what memory agents use. If you build multi-agent systems, design these choices deliberately. The Mindustry environment and the Multi-Agent Harbor framework are open, so you can try your own model mixes.