Four AI agents powered by Google’s Gemini 4 Argon model autonomously beat the game Minecraft for the first time without human assistance, slaying the Ender Dragon in a randomly generated world. According to Google DeepMind, the agents were deployed without preset roles, strategies, or game-specific tools. Over a continuous 27-hour and 14-minute run, the multi-agent system managed task distribution, logged shared memories, generated 2,783 JavaScript programs, sent 229 chat messages, and logged 605 in-game deaths.

DeepMind reported that the total session cost $1,620 based on pricing schedules planned for 2027. The achievement underscores the rapidly expanding capabilities of general-purpose foundation models handling complex, multi-agent coordination problems in open-ended environments without task-specific reinforcement learning.

Why it matters

  • Demonstrates advanced multi-agent coordination, code generation, and shared memory management in complex, open-ended environments.

  • Highlights the transitioning of specialized RL tasks to general-purpose foundation model agents, albeit at substantial inference compute costs.

Source: the-decoder.com