Google DeepMind's Gemini 4 agents defeat Minecraft's Ender Dragon after 605 deaths

Google DeepMind’s Gemini 4 agents defeat Minecraft’s Ender Dragon after 605 deaths

by

Four autonomous artificial intelligence agents powered by Google’s Gemini 4 Argon model have successfully completed Minecraft from scratch in 27 hours and 14 minutes. Operating in a randomized survival world without human intervention, the AI team wrote 2,783 custom programs, coordinated tasks via text chat, and defeated the Ender Dragon.

Google DeepMind has deployed general-purpose coding agents to beat Minecraft with no human help. Unlike historic gaming breakthroughs such as AlphaGo and AlphaStar, which relied on custom systems trained specifically for a single title, the experiment used the identical software assistants engineers deploy for daily programming tasks, as Google DeepMind announced. Google researchers published their paper detailing the experiment on arXiv on October 15.

AI Agents Coordinate Inside a Randomized Survival World

The experiment took place in a randomly generated Minecraft Java Edition 1.20.4 world running in survival mode on Easy difficulty. To start the test, researchers placed a team of four agents into the seed with a single goal: defeat the Ender Dragon.

Each agent controlled a player via an open-source library called Mineflayer and coordinated with its teammates through in-game chat and shared memory.

No roles, strategies, or game-specific tools were assigned ahead of time. Before reaching the final boss, players must harvest and process raw materials, venture into dangerous alternate dimensions to secure rare resources, and use spatial mathematics to locate a hidden underground fortress.

beating the dragon with AI

Agents Write Code and Survive Deaths

Mashable noted that the AI players died 605 times, mostly to monsters, while 22 percent of the written programs failed. A fifth agent called the Observer monitored the session as an independent witness to check for cheating and publish live and summary reports. Over the course of the run, the AI players encountered friction.

IGN detailed that it took around 13.5 hours to reach The End, while the remaining time was spent trying to kill the final boss. Thirty-nine minutes after destroying the final healing crystal, a surviving agent clinging to just 2.5 out of 20 health points struck the decisive final blow against the dragon.

Google DeepMind's Gemini 4 agents defeat Minecraft's Ender Dragon after 605 deaths
Photo: Engadget

Because the AI agents played through code rather than a screen, they accessed information unavailable to human players, such as identifying blocks hidden behind other rocks.

  • Programs written: 2,783 JavaScript programs
  • Chat messages: 229 messages exchanged
  • Deaths: 605 total deaths
  • Token usage: 1.35 billion input tokens

Software Engineering Implications and Pricing

According to The Decoder, the entire session cost $1,620 at the standard Gemini 4 Argon price taking effect in 2027. Google recently opened up Gemini 4 Argon to AI Pro and Ultra subscribers, alongside developers and companies via an API, targeting the model toward long, multi-step tasks including programming.

Four AI Agents Beat Minecraft With Zero Human Help (27 Hours, 605 Deaths)

Researchers emphasized that demonstrating autonomous agents can collaborate, rebound from repeated failures, and stick to a multi-day objective proves they can handle open-ended problem solving required for cutting-edge software engineering, scientific research, and complex physical systems.

The skills required to solve Minecraft such as coordinating a team, recovering from failure, and pursuing a goal for more than a day, are also what agents need to work reliably alongside people.

Gemini 4 Argon Agents Defeat Minecraft’s Ender Dragon Through Autonomous Code Generation.

More on Google: Google Launches Gemini Spark for macOS and Plans New Android Chat Features · Google Launches Public SynthID Portal to Detect AI Watermarks · Google Commits €13B to Fund AI Infrastructure in Finland

Worth a look

Read next