Four AI agents running on Google’s Gemini 4 Argon model defeated Minecraft in a fully autonomous session, slaying the Ender Dragon without any human input. The Google Deepmind team placed the agents into a randomly generated world without assigning specific roles or providing game-specific tools. The four systems split tasks independently, warned one another of threats, recovered items from deceased teammates, and logged errors in a shared memory. The run lasted over 27 hours and 14 minutes. During this time, the agents wrote 2,783 JavaScript programs, exchanged 229 chat messages, and died 605 times. The entire session cost $1,620 at the standard Gemini 4 Argon price scheduled for 2027. This event marks another instance where agentic AI models beat video games that previously required purpose-built systems trained from scratch.
This development matters because it demonstrates that large language models can coordinate complex multi-step goals without explicit programming. The agents effectively managed resource allocation and error recovery through natural language communication. The high cost and long duration highlight the current limitations in efficiency and speed. These results suggest that general-purpose models are narrowing the gap with specialised game-playing AIs.
- Agents wrote 2,783 JavaScript programs
- Session cost $1,620
- Run lasted 27 hours 14 minutes
