YC put 50 agents in Minecraft. The leader was whoever came first in the list
YC Labs and Olam Labs dropped 50 GPT-6.1 Sol agents into a survival Minecraft world for three hours with one goal: "Build a thriving civilization." The agents had a directory of each other, a shared sandbox and direct messages. They sent about 5,900 messages, cleared nearly every tree near spawn within 90 seconds, and ended up with something like a government: a ration ledger, written rules, permits and appointed operators.
How they picked a leader is the part to remember. 46 of the other 49 agents sent their first message to Ada, for no reason except that Ada was "the first name on the roster." Ada became the hub, with 616 DMs received. "There was no vote, election or challenge." Hierarchy came from directory order. Then the first night hit: 22 of the run's 34 deaths happened in it, and the colony stayed cautious for the rest of the run, even though dying cost little more than dropped gear.
The team also ran Bed Wars, ten Claude Opus 5.5 agents against ten GPT-6.1 Sol agents over nine games. GPT won the series 6 to 3, usually by breaking Claude's bed in the last few minutes. Claude played defense and led on kills, by 63 to 16 in one game, figuring it would win on kills if both beds stood. The styles differed sharply. GPT sent 5 to 30 times more messages in clipped shorthand, and its most-messaged player got 198 to 368 DMs a game. Claude signed its name and said "we." In a separate Mindustry experiment, GPT agents fetched the game's public source code when stuck. Claude agents never did.
All of this has logs, recordings and an open-sourced harness, so the claims can be checked. That is rare for multi-agent demos.
Our read: if you run many agents together, the order of your roster, your default hub and your first failure will shape the whole system more than your prompt does. Make leadership and escalation explicit, or the list order will decide for you.
Write-up: https://labs.ycombinator.com/projects/games/index.html Harness: https://github.com/yc-software/minecraft-agent-harness
← Back to all articles
How they picked a leader is the part to remember. 46 of the other 49 agents sent their first message to Ada, for no reason except that Ada was "the first name on the roster." Ada became the hub, with 616 DMs received. "There was no vote, election or challenge." Hierarchy came from directory order. Then the first night hit: 22 of the run's 34 deaths happened in it, and the colony stayed cautious for the rest of the run, even though dying cost little more than dropped gear.
The team also ran Bed Wars, ten Claude Opus 5.5 agents against ten GPT-6.1 Sol agents over nine games. GPT won the series 6 to 3, usually by breaking Claude's bed in the last few minutes. Claude played defense and led on kills, by 63 to 16 in one game, figuring it would win on kills if both beds stood. The styles differed sharply. GPT sent 5 to 30 times more messages in clipped shorthand, and its most-messaged player got 198 to 368 DMs a game. Claude signed its name and said "we." In a separate Mindustry experiment, GPT agents fetched the game's public source code when stuck. Claude agents never did.
All of this has logs, recordings and an open-sourced harness, so the claims can be checked. That is rare for multi-agent demos.
Our read: if you run many agents together, the order of your roster, your default hub and your first failure will shape the whole system more than your prompt does. Make leadership and escalation explicit, or the list order will decide for you.
Write-up: https://labs.ycombinator.com/projects/games/index.html Harness: https://github.com/yc-software/minecraft-agent-harness
Comments