A Google DeepMind team reportedly beat Minecraft without human input using four AI agents powered by Gemini 4 Argon, defeating the Ender Dragon. They played for 27 hours and 14 minutes in a randomly generated world, writing a total of 2,783 JavaScript programs, exchanging 229 chat messages and dying 605 times.
Four Agents Thrown In Without Assigned Roles
What stands out about this experiment is not just the result, but how the agents got there. The team reportedly sent four agents into a randomly generated world without assigning them roles or a strategy, and without giving them game-specific tools. It was, in a sense, like leaving four people on a deserted island without a map or a manual.
The agents divided up the work themselves. They reportedly warned their teammates about threats such as enemies and recovered the belongings of teammates who had fallen. Such behavior comes naturally to human teams, but it is remarkable that AI agents reproduced it without a preset script.
“Autonomy” here probably does not mean operating without any framework. The agents had the Minecraft world as their environment and “defeat the dragon” as their goal; within those constraints, they decided for themselves how to proceed. It is more accurate to see this as an experiment that left the choice of means to the agents, rather than one with zero instructions.
What 605 Deaths Tell Us
The numbers show that this was no straightforward march to victory. Over 27 hours and 14 minutes, the agents wrote 2,783 programs, exchanged 229 messages and died 605 times. That works out to a little over 20 deaths per hour.
What is notable is that the agents did not discard their failures. They reportedly recorded mistakes in shared memory. If one agent fell into a trap, that information became an asset for the whole team. In effect, the AI agents were keeping their own version of the “retrospective notes” used in human workplaces.
The number of programs is also telling. The agents wrote JavaScript to control Minecraft, tested it and revised it. That cycle repeated across 2,783 programs before they finally reached the Ender Dragon. Their success may have owed less to a single brilliant decision than to the persistence to turn each failure into the next move.
$1,620 as a Practical Yardstick
The cost of the session, which lasted more than 27 hours, is said to have been $1,620, calculated using Gemini 4 Argon’s standard pricing, which takes effect in 2027. That amounts to several hundred thousand yen. It may seem like a lot to pay just to beat a game, but it can also be seen as the cost of four agents working for more than a full day without human assistance.
In the past, beating a game required AI trained from scratch specifically for that title. Now, general-purpose models are trying to reach the same level by adapting to their environment as they write programs. Minecraft is one of the latest examples of agentic AI tackling games.
The question is how this kind of coordination could translate to the workplace. When multiple AI agents spend long periods dividing up tasks such as research, development and testing, assigning roles and recording failures will be essential. Will the behavior they demonstrated in a game prove useful in real-world work in a way that justifies the cost? As a test case, the $1,620 figure is worth remembering.
Comments