Four AI Agents Beat Minecraft
This presentation examines a landmark experiment in which four general-purpose coding agents autonomously defeated Minecraft's Ender Dragon after 27 hours, using only the standard Mineflayer library without human-authored routes, skill libraries, or game-specific planning. The talk explores how the agents wrote nearly 3,000 JavaScript programs, developed emergent coordination strategies, survived over 600 deaths, and iteratively diagnosed complex game mechanics to complete thirteen dependent milestones from bare-hand resource gathering to the final dragon encounter.Script
Four AI agents entered a freshly generated Minecraft world with one instruction: defeat the Ender Dragon. No human told them how, where to go, or what to build first. Twenty-seven hours later, they won.
Each agent controlled a Minecraft player by writing JavaScript programs executed through Mineflayer. A program runs for about 30 seconds, returns text describing what happened, then the agent writes the next one. This created a loop of observe, act, and verify repeated nearly 3,000 times.
Within ten minutes, the agents negotiated roles through in-game chat. One gathered food and built beds, another mined diamonds, a third smelted iron, and the fourth prepared the Nether portal. This division of labor was not programmed. It emerged from their conversation and remained stable throughout the run.
Finding the stronghold required geometric problem-solving. One agent threw Eyes of Ender from two locations, sampled their trajectories every 50 milliseconds, then intersected the rays. The calculation landed 43 meters from the actual portal room, demonstrating code synthesis as a tool for spatial reasoning.
The End consumed half the runtime and 69 deaths before one agent diagnosed the problem: ten healing crystals on obsidian pillars kept restoring the dragon to full health. The team spent the next five hours systematically destroying every crystal by stacking dirt and punching them. Thirty-nine minutes after the last crystal fell, the dragon died.
The agents built a library of 42 reusable helper functions, used across 1,500 programs, and corrected each other's false claims through chat. They succeeded not by executing flawlessly, but by converting 605 deaths and a 22 percent program error rate into shared procedural knowledge. If you want to see how autonomous coding agents tackle long-horizon challenges, explore more research like this at EmergentMind dot com.
