Google DeepMind AI agents autonomously complete Minecraft

Researchers from Google DeepMind have demonstrated that four AI agents, powered by the Gemini 4 Argon model, successfully completed Minecraft from start to finish without human intervention. The agents were placed in a randomly generated survival world on easy difficulty with the sole objective of defeating the Ender Dragon.

Over a period of 27 hours and 14 minutes, the agents exhibited emergent collaborative behaviors. Without pre-assigned roles, they organically divided tasks: one agent focused on tool production, another on gathering food and resources, a third on activating portals, and a fourth on constructing shelters and delivering the final blow to the boss. They communicated via in-game chat and a shared memory system to exchange warnings, coordinates, and tips.

The experiment involved significant technical complexity, with the agents writing and executing 2,783 programs and exchanging 229 messages. Despite their success, the agents died 605 times during the run. Researchers noted that while the agents had advantages—such as accessing game data through code rather than a visual screen—the achievement highlights the potential for AI agents to handle complex, multi-step tasks and long-term planning required for real-world software engineering and professional automation.