Four AI agents running Google’s Gemini 4 Argon model beat Minecraft’s final boss, the Ender Dragon, without any human input. Google DeepMind described the run in a new paper, and it took 27 hours and 14 minutes. If you follow AI agents, meaning software that carries out multi-step tasks on its own, this is a notable milestone.
What exactly happened?
The four agents played in a randomly generated Minecraft survival world inside a sandbox. Nobody assigned roles, strategies or game-specific tools. Each agent controlled a player through Mineflayer, an open-source library that lets code drive a Minecraft character. A fifth agent, called the Observer, only watched as an independent witness, checking for cheating and publishing live and summary reports, according to Engadget.
The numbers, per The Decoder and Engadget:
- 2,783 programs written and run by the agents;
- 229 messages exchanged between them;
- 605 deaths in total;
- about 1.35 billion input tokens (tokens are the chunks of text a model reads and writes).
How did they work together?
The agents talked through in-game chat and shared a common memory. On their own, they split up the work and settled into roles. Engadget says one made tools, another gathered food and supplies, another lit the portal, and another built shelters against Endermen, a hostile mob, and landed the final blow.
They also warned each other about threats, passed along tips and coordinates, recovered items from teammates who died and logged mistakes in the shared memory so they wouldn’t repeat them. With 605 deaths over the run, that memory got a lot of use.
How much did it cost?
The Decoder reports the session cost $1,620, calculated at the standard Gemini 4 Argon price that takes effect in 2027. Engadget gives both figures: $810 at Google’s current list price and $1,620 at the standard price. Either way, that’s a lot of money to win a video game, which is why this is still research and not a product.
Why does it matter?
Engadget draws a contrast with AlphaGo, DeepMind’s own system: beating Go took a purpose-built AI. Here, a general model with no game-specific tools finished a long, complex game. The outlet also links a similar GPT-6 Pokémon story as an example of the same trend.
Some caution is in order. The claim that this is the first time agents have completed Minecraft end to end without help comes from the team itself and hasn’t been independently verified, according to Engadget. The test also ran in a single randomized, sandboxed world, and burned through a lot of tokens.
For everyday users, the takeaway is practical. AI assistants that handle long jobs, like planning a trip or fixing a technical problem, rely on the same abilities: planning, dividing work, recovering from failure and remembering what went wrong. Minecraft is a cheap lab for testing that, since mistakes hurt no one, even if the bill is steep. To learn about the model behind the experiment, our coverage of Gemini 4 Argon, price and access explains who can use it and what it does.
Where do we already see agents day to day?
Agents that act on your behalf already show up in consumer products, at a much smaller scale. One example is the ChatGPT Chrome extension, which reads your tabs and carries out tasks in the browser. The difference from the Minecraft experiment is duration and risk: here the agents had more than a day to try, die and try again, and each death only cost time. In real life, a mistake could delete a file or buy the wrong thing, which is why human oversight remains the rule.
The Observer deserves attention too. Having a separate agent whose only job is to check for cheating is an idea likely to repeat: the more autonomy an AI gets, the more it matters to have an independent check on what it did.
