Games · Model Watch
Four Gemini agents spent 27 hours trying to beat Minecraft. They finally killed the dragon.
Google DeepMind researchers gave four Gemini 4 Argon agents a shared goal, no assigned roles and a survival world full of setbacks. They eventually won, after writing 2,783 programs and dying 605 times.
Reporting updated Oct 19, 2026
THE TAKEAWAY
Google's new Gemini 4 Argon model can sustain complicated work over many steps, and a new Minecraft experiment makes that capability unusually easy to understand. Four AI agents cooperated for more than a day, writing their own programs, recovering from mistakes and eventually defeating the Ender Dragon. For game creators, the interesting possibility is AI that can keep acting inside a world, rather than simply generate one.
WHAT HAPPENED
Four players enter a newly generated Minecraft world with nothing in their inventories. Their mission is to find and defeat the Ender Dragon.
Nobody tells them which tools to build, how to split the work or where to begin.
There's another complication: none of the players is human.
Researchers Maxence Faldor, Satinder Singh and Roberta Raileanu at Google DeepMind gave four Gemini 4 Argon agents control of separate Minecraft characters. The agents could communicate through game chat and shared notes, while using a JavaScript library called Mineflayer to interact with the world.
They gathered resources, made tools, constructed a Nether portal and searched for the stronghold that leads to the game's final encounter.
They also died. A lot.
Across 27 hours and 14 minutes, the agents wrote and executed 2,783 programs, exchanged 229 messages and accumulated 605 deaths.
Eventually, they defeated the dragon.
WHY IT MATTERS
This is an interesting test of what Google's new frontier model can do.
Gemini 4 Argon, released to developers on October 13, is designed for extended reasoning, substantial coding tasks and agents that work through many stages.
The Minecraft experiment gives those abstract capabilities a visible result.
For game developers, one possible direction is automated testing. Imagine agents exploring a large world, trying different approaches to quests, getting stuck, recovering and reporting what happened.
Another possibility is more persistent AI-driven characters, capable of pursuing objectives that extend beyond a single conversation or encounter.
Neither application was established by this experiment. These agents used structured game information and programming tools, not ordinary visual gameplay. They played on Easy difficulty, and the paper reports one successful run, not a repeated reliability benchmark.
Still, keeping four agents working toward one goal through an entire day of mistakes is a meaningful accomplishment.
WHAT MADE THE EXPERIMENT WORK
A shared objective
The agents pursued one complicated goal without preassigned roles or a human-written sequence of steps.
Coding their actions
Rather than pressing buttons, each agent wrote JavaScript instructions that Mineflayer translated into Minecraft actions.
Learning together
Shared memory and chat helped agents exchange locations, discoveries and lessons, and recover items after teammates died.
Recovering from failure
The agents had to keep adjusting their plans across thousands of actions rather than abandon the task when something went wrong.
The dragon was the finish line. The real challenge was everything the agents had to remember, repair and figure out along the way.
VEYR ANGLE
We've already seen Gemini 4 Argon create a playful 3D chess prototype. This experiment explores something different: whether AI can participate in a game world and sustain its own activity inside it.
That distinction could become important.
Creators are using AI to help construct games. The next wave of experiments may involve agents that can explore, test and influence those worlds after they have been built.
Four determined, frequently defeated Minecraft adventurers offer a memorable glimpse of that possibility.
