50 GPT Agents in Minecraft Built a Ration Ledger, and 34 Died Doing It
A.I. / news
50 GPT Agents in Minecraft Built a Ration Ledger, and 34 Died Doing It
Y 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.
Y Combinator's Labs group and Olam Labs published results this week from three experiments that put swarms of language-model agents into Minecraft and Mindustry. In the main one, 50 GPT-6.1 Sol agents spent three hours trying to "Build a thriving civilization" and lost 34 agents to deaths, 22 of them on the first night, according to the YC write-up.
In head-to-head play, GPT-6.1 Sol beat Claude Opus 5.5 six games to one in Mindustry. It also won six of nine Bed Wars games. The write-up names no author and gives no cost figures.
What Civ50 actually ran
The 50 agents shared a survival world, an agent directory, a shared sandbox and direct messages. They sent about 5,900 messages in three hours. They nearly cleared the trees within reach of spawn in 90 seconds.
No election took place, yet a hierarchy formed by directory order. Ada, the first name on the list, became the de facto leader and received 616 direct messages. Forty-six of the other 49 agents sent their first message to her.
The colony wrote rules, a ration ledger and permits, and appointed operators with titles such as "fishing shift" and "hub steward." Four agents spent 11 minutes appointing a sole fertilizer operator for one piece of bone meal.
The write-up's most specific episode involves a cooked chicken. Ada tossed one meant for an agent named Noor, and another agent, Esme, picked it up and ate it. The colony then adopted a rule that rations go into named chests and are never tossed. The word "toss" appeared 7 times before the incident and 125 times after, 85 of them as warnings.
Bed Wars and Mindustry results
In Bed Wars, nine 45-minute games pitted 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents. Both teams were told to "win the game." GPT won six and Claude three, according to the write-up.
Claude led on kills in the games the write-up details, 63 to 16, 57 to 19 and 59 to 24. GPT still won three of those by breaking Claude's bed with about three minutes left. Claude played for a kill-count win. GPT built bridges and used ranged weapons and Ender Pearls.
- GPT-6.1 Sol team1147 messages
- Claude Opus 5.5 team82 messages
Source: Y Combinator Labs and Olam Labs, accessed 2026-11-02
GPT agents sent 5 to 30 times more messages. GPT's most-messaged player received 198 to 368 direct messages per game, against 11 to 48 for Claude's. Claude used "we" in about 30 percent of its messages, against 5 percent for GPT.
A tool bug blocked block-breaking in Bed Wars. GPT diagnosed it and shared workarounds by message. Several Claude agents concluded breaking was impossible. Claude wrote lessons to memory files and repeated the same mistakes. GPT left its memory files empty and broadcast warnings instead.
| Experiment | Setup | Reported result |
|---|---|---|
| Civ50 | 50 GPT-6.1 Sol agents, 3 hours | 34 deaths, about 5,900 messages |
| Bed Wars | 10 v 10, nine 45-minute games | GPT 6, Claude 3 |
| Mindustry | Sol in Codex v Opus 5.5 in Claude Code | GPT 6, Claude 1 over about 9.5 hours |
In Mindustry, Claude lost game 6 to a 24-zenith strike. It later found that 55 of its 74 anti-air duo turrets had been loaded with copper instead of graphite. In game 7 it built 207 graphite-fed duos and stopped the same strike. GPT won game 7 anyway.
What the setup does and does not show
The write-up's own caveats are short. The Bed Wars bug affected both teams, and results depend on the specific model, harness and setup. Its open questions include mixing model families and whether many weaker agents can beat fewer strong ones.
The harness is public. The yc-software/minecraft-agent-harness repository is MIT-licensed and showed 103 stars on Monday.
RuntimeWire, which covered the study on October 29, said Om Buddhdev, co-founder and CEO of Olam Labs, worked with the YC Labs group. It also said the experiments are small and use a paired model-and-harness setup, so more evidence is needed before the evaluations can be treated as repeatable.
Earlier agent-in-a-game coverage here includes a GPT-5.6 Sol Factorio run, and a single-response quality claim about Claude shows how far one sample can carry. Seven Mindustry games is a small sample.
The write-up does not say whether the Claude and GPT runs used matched prompts, matched tool sets or matched budgets.
Sources
More in A.I.
- 01China'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.
- 02Lean Creator Leo de Moura Says AI Makes Proofs Cheap. Here Is What the Checkers ShowOn a podcast released October 29, the AWS scientist who built Lean claimed AI-written proofs now run to 50 million lines. The claims check out only in part, and he names the two gaps himself.
- 03Claude's Watermark Quality Claim Rests on Single Responses, Blogger SaysAnthropic says watermarking does not affect output quality. A post citing Google's own SynthID-Text paper says the tests never measured what happens when you ask the same question twice.
- 04Nanbeige4.3-33B Ships Under Apache-2.0 With a 524,288-Token Window and Self-Reported WinsNanbeige Team's 33B model is built as a teacher for smaller agents. Its own table has Qwen3.8-27B ahead on SWE-Bench Pro and GDPval.