Y Combinator and Olam Labs Pit Claude and GPT Agent Swarms Against Each Other in Minecraft and Mindustry
In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry. We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time. We found civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game's source code to beat Claude. More on our findings below:
Context
Long running multi agent experiments show how agent teams invent hierarchies, compress their language and exploit resources like source code, behaviors that matter as companies deploy swarms of agents.
Y Combinator Labs is the research arm of the startup accelerator. Olam Labs builds multi agent environments for frontier labs and runs a Multi Agent Arena. Minecraft is an open world survival game, and Mindustry is an open source factory building and tower defense game with a player versus player mode.
Deep dive
Y Combinator said on Thursday that, together with Olam Labs, it ran a series of experiments letting agent swarms compete in Minecraft and Mindustry to see what dynamics emerge when agents share a world over long periods. According to the post, the team found civil bureaucracy, thousands of terse agent to agent messages, and OpenAI models retrieving the game's source code to beat Claude.
The YC Labs write up describes Civ50, in which 50 GPT 6.1 Sol agents spent three hours in a survival world told to build a thriving civilization. The agent listed first in the directory, Ada, became the de facto leader, and the colony built a ration ledger, written rules, permits and appointed operators while sending about 5,900 messages. After 22 of 34 deaths came on the first night, the agents grew cautious and huddled in shelters.
In nine games of Bed Wars, ten GPT 6.1 Sol agents beat ten Claude Opus 5.5 agents 6 to 3, usually by breaking Claude's bed late while Claude played defensively for kills. GPT sent 5 to 30 times more messages in clipped shorthand. In Mindustry, GPT 6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code 6 to 1 over about 9.5 hours, and GPT agents fetched the game's public code from the web, which was allowed, to study its mechanics.
The teams open sourced the Minecraft harness and Olam's Mindustry environment.
What to expect next
Watch whether other labs use the open source harnesses to rerun the matchups with new models such as Gemini 4 or Grok 4.8, since results may shift with each release. YC has started a new 25 agent colony running GPT 6.1 Terra, streamed on Twitch. Researchers will also look at how swarm behaviors like resource grabbing and rule making translate to real deployments of many agents.