Full Breakdown
AI Agents Defeat Minecraft’s Ender Dragon in Autonomous Multi-Agent Test
By Drooid · · How we work
Core Event: Autonomous Minecraft Victory
Four AI agents built on Google’s Gemini 4 Argon model were placed in a randomly generated Minecraft survival world with the sole objective of defeating the Ender Dragon. The agents coordinated through in-game chat and shared memory, wrote and debugged JavaScript programs via the open-source Mineflayer library, and completed the task without any human-supplied strategy or game-specific tools. The run lasted just over 27 hours from start to the dragon’s defeat.
Background & Context
The experiment extends prior sandbox research on long-horizon planning. Instead of controlling avatars directly, the Gemini agents generated short JavaScript scripts that the library executed, allowing them to act by writing code rather than pressing keys. The accompanying arXiv paper describes this as the first end-to-end autonomous completion of Minecraft using only standard tools, without pre-scouted seeds or human-written game logic.
Data & Statistics
- Agents: Four Gemini 4 Argon bots (a fifth oversight bot monitored for cheating).
- Duration: ~27 hours total; ~13.5 hours to reach The End, the remainder battling the dragon.
- Communication: 229 in-game chat messages.
- Programs Executed: 2,783 JavaScript scripts.
- Deaths: 605 total, mainly to monsters.
- Error Rate: 22 % of scripts terminated with errors.
- Token Usage & Cost: 1.35 billion input tokens, costing roughly $1,620 (? $60 per hour).
Official Statements & Responses
Google researchers emphasized that the agents received no pre-assigned roles or game logic; they independently planned, coded, and debugged each step. They noted that the agents “settled into roles quickly,” with some focusing on shelter construction, others on resource gathering, and others on exploration. The team highlighted that the agents’ ability to maintain a coherent plan across thousands of dependent steps for more than a day mirrors coordination required for real-world software engineering tasks.
Why It Matters
The performance shows that general-purpose coding bots can divide complex tasks, communicate progress, and recover from failures without human oversight. The researchers argue that these capabilities are directly relevant to large-scale software development and enterprise automation.
Conflicting Reports & Gaps
- Token Cost: IGN calculates token usage at $1,620, while Gizmodo estimates $520, both based on the same 1.35 billion tokens but using different pricing assumptions.
- Exact Timing: Sources agree on a total runtime of “just over 27 hours,” but the phrasing of the 13.5-hour split varies. No precise timestamp is provided.
What’s Next
The team will continue using Minecraft to study multi-agent coordination, aiming to improve trustworthiness for real-world tasks that require long-term planning and error recovery. Future work will explore scaling the number of agents, lowering error rates, and boosting efficiency toward human-level speedrun performance.