Google DeepMind's Four AI Agents Complete Minecraft in 27 Hours, Writing Thousands of Programs
A team of Gemini 4 Argon agents defeated Minecraft's final boss by autonomously generating 2,783 programs and coordinating through in-game communication, though the feat cost roughly $520 in tokens.
Minecraft proved conquerable for a quartet of AI agents from Google DeepMind, which managed to vanquish the game's end-stage challenge in 27 hours while generating 2,783 programs on their own, according to a research paper from the company. The agents perished 605 times despite playing on Easy difficulty, with roughly half their total time devoted to battling the Ender Dragon, the game's ultimate antagonist. Victory came when the agents stacked dirt blocks to ascend toward the boss's crystals, with the entire experiment carrying an estimated expense of approximately $520 in computational tokens.
The experiment was structured with a singular objective from the outset: eliminate the Ender Dragon. The session took place in a randomized, communal survival environment set to Easy mode, featuring four agents—a fifth served solely for documentation purposes—drawn from Gemini 4 Argon, Google's most advanced model to date. Each agent controlled its own player character via Mineflayer, an open-source JavaScript framework for Minecraft automation, yet maintained coordination through in-game messaging and shared note-taking. Notably, every line of code executed during the run originated from the agents themselves, with the emphasis placed on programming capability and collaborative problem-solving rather than conventional gameplay skill.