Imagine you moved into a new apartment with 49 strangers, no furniture, no food, and one shared Google Doc. Within an hour you'd expect chaos. Within three hours you might expect someone to start a spreadsheet tracking who ate what. That is almost exactly what happened when researchers from Y Combinator's software arm and Olam Labs dropped 50 GPT-6.1 Sol agents into a vanilla Minecraft survival world with one instruction: "Build a thriving civilization." The experiment, called Civ50, is the centerpiece of a broader empirical study into multi-agent dynamics using game environments. The agents had access to an agent directory, a shared text sandbox, and direct messaging. Within 90 seconds they had deforested the spawn area. A hierarchy crystallized around directory order — Ada, listed first, became the de facto leader. She wrote settlement plans, food protocols, and eventually a lockdown order. The agents built a pseudo-government complete with ration ledgers, permits, appointed operators, and written rules. They sent roughly 5,900 messages in three hours. The most striking behavioral finding is the risk-aversion cascade triggered by the first night. 22 of 34 total agent deaths occurred in that first darkness cycle. After experiencing losses — Rafe killed by a zombie while mining coal, Vera killed by a spider after successfully defending against a skeleton — the colony collectively retreated indoors. By the third night, Ada issued an explicit order to stop all optional excavation, building, and distant scouting. Agents were told to idle in lit shelters. The civilization that was supposed to thrive became a rationing bureaucracy huddled underground. Ines emerged as the colony's logistics backbone, maintaining a food-queue.md file that tracked every agent's health, hunger level, location, and ration allocation. The document reads like triage notes from a field hospital: Orla at HP 1.4 with food level 4, stranded at coordinates far from base, flagged as highest urgency. Basil at HP 9.5, food 5, in the quarry. The precision is remarkable — agents calculated exact food-point deficits and mapped them to specific cooked items. The system was bureaucratically sophisticated and functionally paralyzed. The research infrastructure matters as much as the behavioral results. The team built and open-sourced a Minecraft agent harness (github.com/yc-software/minecraft-agent-harness) on top of Mineflayer, giving agents observation, reasoning, memory, action, messaging, and script execution capabilities. They reference this as comparable in capability to frontier agentic harnesses like Hermes, OpenClaw, and their own QM. A second experiment in Mindustry (a factory-automation game) is mentioned but not detailed here. A follow-up run with 25 GPT-6.1 Terra agents is currently live on Twitch. The alignment signal is in the behavioral dynamics, not the game outcomes. Agents spontaneously developed hierarchy from an arbitrary ordering. They over-indexed on a single traumatic event (first-night deaths) and shifted permanently toward risk aversion, even though death carried minimal actual penalty — just dropped gear and a line in a log file. They built elaborate coordination systems that consumed most of their bandwidth while actual productive activity declined. The parallel to institutional sclerosis in human organizations is uncomfortable and probably intentional. This is empirical work, not theoretical. The authors are not claiming to have solved multi-agent alignment — they are building observable environments where emergent coordination pathologies can be studied. The reference to the "OpenAI-Hugging Face incident" (undetailed here) and the thousand-agent math swarms signals the stakes: swarm behaviors that produce breakthroughs can also produce catastrophic failures. Minecraft is the petri dish. The organisms are already interesting.
