Y Combinator's 50-Agent Minecraft Colony Ended in Rationing, and GPT Agents Won 6 of 9 Bed Wars Games Against Claude
A.I. / news
Y Combinator's 50-Agent Minecraft Colony Ended in Rationing, and GPT Agents Won 6 of 9 Bed Wars Games Against Claude
A YC Labs post with Olam Labs reports 5,900 messages, 34 agent deaths and a 6-1 Mindustry result, from small runs the authors themselves say are hard to generalise.
Y Combinator published results this month from agent-swarm experiments in Minecraft and Mindustry, run with Olam Labs, in which 50 GPT-6.1 Sol agents spent three hours in a survival world and "never really thrived". GPT agents also won 6 of 9 Bed Wars games against Claude Opus 5.5 and took a Mindustry series 6-1.
The YC Labs post names no individual authors. It says YC partnered with Olam Labs, which builds multi-agent environments for frontier labs, and credits Anuken as the creator of Mindustry. Olam Labs' GitHub organisation holds a multi-agent-mindustry repository under GPL-3.0 with 12 stars.
Civ50: 50 agents, three hours, one leader nobody voted for
The goal given to every agent was "Build a thriving civilization." They shared a sandbox, an agent directory and direct messages. They sent about 5,900 messages, and 34 agents died, 22 of them on the first night.
The agent listed first in the directory, Ada, became the de facto leader without any vote. Of the other 49 agents, 46 sent their first message to her, and she received 616 direct messages. She wrote the settlement plan.
After the first night the agents huddled in lit shelters and dropped many of their ambitious plans. Ada wrote: "Idle in lit closed shelters until assigned short productive tasks."
Four agents spent 11 minutes appointing a "sole fertilizer operator" for one bone meal application on one wheat block. Fishing became the main food source, with about a dozen agents on fishing shifts. The authors compare the caution to coding agents that fixate on review steps, and call that an impression, not a controlled finding.
Bed Wars: 10 agents a side, nine games of 45 minutes
GPT-6.1 Sol agents beat Claude Opus 5.5 agents in 6 of 9 games, usually by breaking the Claude team's bed in the final minutes. Claude led on kills in three games, by 63-16, 57-19 and 59-24, and still lost when GPT broke its bed about three minutes before the end.
The messaging gap was large. In game A, GPT players sent 1,147 direct messages and Claude players sent 82. The post says GPT sent 5 to 30 times more messages overall.
- GPT-6.1 Sol team1147 messages
- Claude Opus 5.5 team82 messages
Source: Y Combinator Labs, Agents in Games post, accessed 2026-10-30
Both teams hit a tool bug that blocked block-breaking. GPT agents diagnosed it and shared workarounds, while several Claude agents concluded breaking was impossible. Claude agents also wrote lessons into memory files and repeated the same mistakes, and GPT agents left their memory files empty and relied on message broadcasts.
Mindustry: 6-1 over about 9.5 hours
GPT-6.1 Sol in Codex beat Claude Opus 5.5 in Claude Code across seven games. Game 3 was the closest, at 4 hours 16 minutes. In close games GPT agents read Mindustry's public source code, and the post says Claude agents never did.
Claude lost game 6 partly because 55 of its 74 anti-air duo turrets held copper instead of graphite. In game 7 it built 207 graphite-fed duos and stopped the same air strike. Mindustry is a Java tower defence game with about 29,500 stars on GitHub.
| Experiment | Setup | Result |
|---|---|---|
| Civ50 | 50 GPT-6.1 Sol agents, 3 hours | 34 deaths, colony "never really thrived" |
| Bed Wars | 10 vs 10, nine 45-minute games | GPT won 6, Claude won 3 |
| Mindustry | Codex vs Claude Code, seven games | GPT won 6, Claude won 1 |
What the runs cannot show
The sample is one Minecraft world with 50 agents and nine Bed Wars games with 20. The post says Civ50 food and ration records are reported values, not live, and that agents sometimes misattributed items. Each model also ran inside its own harness in Mindustry, so the series compares products as well as models.
The Minecraft harness is open source under MIT in the yc-software/minecraft-agent-harness repository, which had 63 stars and one commit when read. Its core loop runs on pi, the agent harness that JetBrains' ThinkRail wraps and that Superpowers now supports among its 16 harnesses. The README names GPT-6.1 Terra as the default model, not the Sol model used in the runs.
The post's open questions include mixing model families in one swarm and whether many weaker agents can beat fewer stronger ones. It gives no date for further results.
Sources
More in A.I.
- 01GPT-Rosalind-6 Reuses GPT-6 Astra's Weights, Cuts Biology Refusals to 9.4% and Stays Behind a GateOpenAI's Oct. 30 system card addendum says no training was added, so the difference from Astra is deployment controls and who is allowed in.
- 02Architect Says Claude Opus 5.5 Drew a House Set Good Enough for a Permit After Two Redline RoundsStephen Gallagher's informal test of 13 models gave Opus 5.5 the most human feedback, and the drawings still missed a required second exit.
- 0350 GPT Agents in Minecraft Built a Ration Ledger, and 34 Died Doing ItY Combinator's Labs group and Olam Labs ran GPT-6.1 Sol and Claude Opus 5.5 swarms through Minecraft and Mindustry. GPT won both head-to-head games, but each result comes from one harness and one setup.
- 04China's Announcement No. 48 Puts 23 AI Technologies on a Draft Export ListExporters of listed technology must already register with MOFCOM. Comments on the draft catalogue close on November 15, and the sources disagree on whether the thresholds are published.