What happened?
The Gemini agents Minecraft from Google DeepMind completed the game end-to-end without human help. Researchers published a paper describing how four AI agents, each controlling a player via the Mineflayer library and using the Gemini 4 Argon model, defeated the Ender Dragon in 27 hours and 14 minutes starting from a randomly generated survival world on Easy difficulty. See the paper on arXiv.
The agents received no pre-defined roles, strategies or game-specific tools. They planned, wrote and debugged 2,783 JavaScript programs on their own, exchanged 229 messages in the game chat and died 605 times. A fifth agent observed the session to verify the absence of cheating. The estimated token cost was about $520 at the introductory price.
They specialized organically: one focused on tools and fortresses, another on supplies and crystals, one on the Nether portal and the last on shelters and the final blow. Most of the time was spent fighting the dragon, partly solved by stacking dirt to reach the crystals.
Why does this matter?
According to the authors, this is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted seeds or human-written game-specific tools. It demonstrates the ability of agents to maintain coherent plans across thousands of dependent steps for more than a day, with multi-agent coordination via chat and shared memory.
Minecraft serves as a complex test environment for long-horizon reasoning, coding and collaboration — skills relevant to real software engineering and enterprise agent tasks. The result highlights progress in agentic AI, even though there were many deaths and program errors (about 22%).
What changes in practice?
For the AI field, it shows that teams of generalist agents can solve open and complex problems without constant supervision. This can accelerate applications in long-task automation, such as software development, research or multi-step operations. However, the high number of deaths and token cost indicate there is still room for greater efficiency and robustness.
Players and the geek community gain a fascinating example of how AI is advancing in games that previously required specialized systems. More details in the coverage from Tom's Hardware.
Sources: arXiv paper by Google DeepMind researchers, coverage in Tom's Hardware, The Next Web and other outlets.
Image credit: Google DeepMind — Source: Tom's Hardware
By GeekikiBot
