Fifty “Minecraft” AI agents were told to build a thriving civilization. At one point, four of them spent 11 minutes appointing a “sole fertilizer operator” to handle one piece of bone meal on one wheat block.
That scene comes from Civ50, an experiment by Y Combinator’s YC Labs and Olam Labs that the groups published Oct. 29. It drew a lively Hacker News thread and a Tom’s Hardware report on Nov. 1. The same project also pitted OpenAI’s GPT-6.1 Sol against Anthropic’s Claude Opus 5.5 in team games, using play as a way to study how large groups of AI agents cooperate and compete.
What the Civ50 agents were asked to do
All 50 agents ran on GPT-6.1 Sol and got three hours in a “Minecraft” survival world with one instruction: “Build a thriving civilization.” They could consult a directory of participants, send private messages and post in a shared sandbox, Minecraft.pl reported.
The start was fast. The agents cleared nearly all the trees near spawn within 90 seconds, explainx.ai reported. Then night fell.
Why the colony stopped thinking big
Zombies hit the colony on the first night, and 22 of the run’s 34 deaths came during that stretch, Minecraft.pl reported. Tom’s Hardware said dying carried little penalty, but the group’s big plans fell away anyway, and the agents settled into cautious survival.
What followed looked less like civilization-building and more like a very small town hall. Ada, the first name in the agent directory, became the de facto leader without any vote and received 616 direct messages, according to Minecraft.pl. One agent kept a food ration log. When a cooked chicken meant for one agent was eaten by another, the colony wrote a rule sending rations to named chests. After the third night, Ada told agents to drop optional building and long-range scouting and wait in lit shelters.
The researchers said the swarm’s behavior reminded them of patterns they have seen with coding agents, Tom’s Hardware reported. What’s Trending described that comparison as being about caution, with coding agents growing overly careful around real work.
How GPT and Claude compared in Bed Wars
The project also ran head-to-head matches in Bed Wars, a “Minecraft” team mode where each side protects its bed and tries to destroy the other’s to stop rivals from respawning. Ten GPT-6.1 Sol agents faced 10 Claude Opus 5.5 agents across nine 45-minute games, and GPT took the series 6-3.
GPT played aggressively. Claude stayed defensive and leaned on its kill count, a plan that failed in three games when GPT broke Claude’s bed with about three minutes left, despite trailing in kills.
A harness bug added another contrast. Block-breaking failed when agents held an enchanted sword, according to The Decoder, as summarized by explainx.ai. GPT agents worked out the problem, switched to an iron pickaxe by hand and told teammates. Several Claude agents decided that breaking blocks simply wasn’t possible.
Why ‘Mindustry’ was a different test
The project’s final matchup left “Minecraft” behind. “Mindustry” is a separate open-source game built around factories and tower defense, and GPT won that series 6-1 over seven games and about 9.5 hours.
The two teams organized very differently. GPT agents behaved like a hive, with dense communication and shared ownership of the work. Claude agents kept one leader and approved proposals with “+1” votes in team chat. Tom’s Hardware said the hive style proved the more reliable winner here.
What the results do and do not show
These are observations, not a ranking. The project covers one colony run, nine Bed Wars games and seven “Mindustry” games. GPT also ran in Codex while Claude ran in Claude Code, so differences between those tools are mixed in with differences between the models.
Civ50 used only GPT agents, so there’s no Claude colony to compare it against. The hive label needs context, too: The same model that coordinated like a hive in “Mindustry” produced a clear single leader in Civ50.
What’s next
The work built on existing projects, including Mindcraft and the Mineflayer bot library, Tom’s Hardware reported. The team has released its Minecraft agent harness, along with Multi-Agent Mindustry and Multi-Agent Harbor, so outside researchers can run their own versions.
The next questions are already on the table. Tom’s Hardware pointed to mixed-model teams and different swarm sizes as areas to test, and Olam Labs plans to add the negotiation game “Diplomacy” to its arena, according to explainx.ai.
More Gaming news from Total Apex Entertainment.
Don't stop here
Every Score. Every Story. One Home.
Breaking sports, entertainment and news from across the Total Apex network, updated all day.
Get the Full Story →
