Google DeepMind AI Agents Beat Minecraft in 27 Hours Using Self-Written Code
Four Gemini 4 Argon agents autonomously defeated Minecraft’s final boss by writing thousands of programs and coordinating via chat without human help.
Key points
- Four AI agents completed Minecraft in 27 hours and 14 minutes using 2,783 self-written programs.
- The team coordinated roles for resources and boss fights through in-game chat without human assignment.
- Researchers claim this is the first end-to-end autonomous completion using Mineflayer without human game logic.
Google DeepMind researchers announced that four Gemini 4 Argon agents successfully defeated Minecraft in 27 hours and 14 minutes. According to a paper published by the company, the agents wrote 2,783 programs entirely on their own to achieve this milestone. The run required no human input during the execution phase, marking a significant step in autonomous AI coordination.
The experimental playthrough took place on a random, shared survival world set to Easy difficulty. The team consisted of four primary agents, with a fifth agent dedicated solely to reporting outcomes. Each agent controlled a player character using the open-source Mineflayer JavaScript library. They coordinated their actions through in-game chat and shared notes, demonstrating complex teamwork without pre-assigned roles or human-written game logic.
Autonomous Coordination and Strategy
Within ten minutes of starting, the agents proposed a division of labor through chat. Bot janesick suggested the initial plan, which was refined by Bot Auratian and acknowledged by Bot Gyldenia. The team assigned specific responsibilities: Gyldenia handled food, beds, and flint; Bulicame focused on diamonds; Janesick managed iron and tools; and Auratian constructed the Nether portal. This self-organized structure allowed them to progress efficiently toward the game’s final objective.
The agents faced significant challenges, dying 605 times during the run. Approximately half of the total time was spent fighting the Ender Dragon, the game’s final boss. The run concluded when the agents stacked dirt to reach the boss’s crystals, a strategy that cost an estimated $520 in tokens. Researchers noted that the agents diagnosed issues, such as Ghast fireballs destroying items, and wrote new rules to prevent future losses.
Technical Implementation and Claims
Google DeepMind describes this achievement as the first end-to-end autonomous completion of Minecraft by AI agents, to their knowledge. This claim is qualified by the use of Mineflayer without pre-scouted seeds or other human-written game-specific tools. The library provided pathfinding and block detection, but all higher-level logic was generated by the agents. Finding the stronghold required janesick to triangulate locations by sampling Eye of Ender flight data every 50 milliseconds.
One agent, Auratian, drowned while carrying eyes, but the team recovered the items and noted the location. The researchers emphasized that no one assigned roles or wrote game logic during the run. The focus of the experiment was on coding and teamwork capabilities rather than traditional game-playing skills. The successful completion demonstrates the ability of frontier models to solve complex, long-horizon tasks through autonomous programming.
What should you do?
Security and technology professionals should monitor developments in autonomous AI agents and their potential for unsupervised execution. Organizations deploying AI systems should implement strict oversight for any agents capable of writing or modifying code. It is important to review how AI models coordinate actions and handle errors in complex environments.
What to do and how to stay safe: Google DeepMind
- Monitor vendor announcements regarding autonomous AI capabilities and potential security implications.
- Implement strict oversight for any AI systems capable of writing or modifying code without human review.
- Review how AI models coordinate actions and handle errors in complex, multi-step environments.
- Ensure that AI agents operate within defined boundaries and cannot execute unauthorized programs.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How long did it take the AI agents to beat Minecraft?
The four Gemini 4 Argon agents completed the game in 27 hours and 14 minutes, according to Google DeepMind.
Did humans help the agents during the run?
No, there was no human input during the run, and no one assigned roles or wrote game logic.
What tool did the agents use to control their characters?
The agents used the open-source Mineflayer JavaScript library for Minecraft bots to control their players.
