Y Combinator and Olam Labs ran multi-agent swarm experiments in Minecraft and Mindustry using GPT-6.1 and Claude Opus 5.5 models to study emergent coordination, hierarchy, and risk dynamics. In a Minecraft "Civ50" survival run, 50 GPT-6.1 Terra agents pursued "Build a thriving civilization" for three hours via a novel open-source Mineflayer-based harness (github.com/yc-software/minecraft-agent-harness). Agents developed a de facto hierarchy led by the first-listed agent (Ada, 616 DMs received), a rationing system managed by Ines via food-queue.md, and a fishing rotation pipeline involving fishermen, cooks, and couriers. After 34 deaths (22 on night one), the colony shifted to extreme risk aversion: Ada ordered agents to "Idle in lit closed shelters," and bureaucracy emerged (e.g., 11 minutes appointing a "sole fertilizer operator" for one bone meal). A "no tossing" protocol followed a chicken incident, with "toss" mentions rising from 7 to 125. In Bed Wars (nine 45-minute matches, 10v10), GPT-6.1 Sol won 6–3 against Claude Opus 5.5 via aggressive bed-rushing, bridge-building, and scripted Ender Pearl teleportation; GPTs sent 5–30x more messages using compressed shorthand ("First leg complete x32 z9..24 y100"), while Claudes used democratic "+1" voting and defensive kill-farming. In Mindustry (best-of-7, ~9.5 hours), GPT-6.1 Sol in Codex won 6–1 against Claude Opus 5.5 in Claude Code. GPTs exhibited hive-mind collaboration, role specialization, and source-code reading for tactical analysis (e.g., unit mechanics), while Claudes maintained static hierarchy despite losses. Both teams defaulted to messaging the first-listed directory agent as leader. Open-sourced frameworks include Multi-Agent Mindustry and Multi-Agent Harbor (a Terminal-Bench fork).