Gemini 4 Argon Agents Defeat Minecraft Autonomously
Google DeepMind's Gemini 4 Argon agents defeated Minecraft's Ender Dragon entirely on their own, demonstrating advanced multi-agent cooperation and complex problem-solving in open-world environments.
A team of four artificial intelligence agents running on Google DeepMind's Gemini 4 Argon model has successfully beaten the video game Minecraft without any human assistance or intervention. Dropped into a randomly generated world with zero human input, the system operated entirely autonomously. The team was not given any game-specific tools, pre-assigned operational roles, or explicit strategies, yet they organized themselves to navigate the survival title's open-ended environment and ultimately slay the Ender Dragon.
Throughout the run, the four agents exhibited flexible division of labor and active peer communication. They routinely split up tasks based on immediate needs, called out dangerous threats to protect one another, retrieved inventory dropped by dead teammates, and tracked strategic errors by logging mistakes into a shared memory buffer. Over the course of 27 hours and 14 minutes, the agentic collective created a total of 2,783 JavaScript programs, sent 229 inter-agent chat messages, and experienced 605 in-game deaths before securing victory.
The computational session incurred a total cost of $1,620 based on the standard pricing model for Gemini 4 Argon scheduled to go into effect in 2027. The achievement provides a striking demonstration of how multi-agent workflows are shifting practical AI development. Until recently, mastering open-ended environments like Minecraft required researchers to design specialized, purpose-built AI systems trained from scratch. Modern frontier agentic frameworks now show that general-purpose foundation models can handle long-context reasoning, programmatic action generation, and autonomous teamwork directly.
This is our own summary of reporting by The Decoder
