Four AI agents running on Google’s Gemini 4 Argon completed Minecraft without human input, ending the run by defeating the Ender Dragon. Their effort relied on the agents coordinating with one another in a randomly generated world, rather than following a human-assigned plan.
Agents divided the work themselves
The Google Deepmind team placed the four agents in the world without assigning roles or strategies. They also received no game-specific tools. The agents had to decide how to approach the challenge as it unfolded.
They split tasks among themselves and warned teammates about threats. When an agent died, others recovered its items. They also kept a shared memory where they recorded mistakes, giving the group a way to retain lessons from earlier attempts.
This kind of coordination was central to the run. The agents were not simply acting independently: their chat messages, warnings, item recovery, and shared notes helped them respond as a team.
A long run with repeated setbacks
The session lasted 27 hours and 14 minutes. During that time, the agents wrote 2,783 JavaScript programs, exchanged 229 chat messages, and died 605 times before completing the game.
Those figures show the amount of trial and error involved. The agents’ success came after a prolonged sequence of actions and failures, with their shared memory and communication supporting further attempts. The source does not describe each individual step in the route to the Ender Dragon, but it does make clear that the run required sustained autonomous play.
What the Minecraft result demonstrates
The achievement is presented as another example of agentic AI models beating video games. The article contrasts these models with game-playing systems that, just a few years ago, required purpose-built AI trained from scratch.
In this case, the agents were given neither designated roles nor a ready-made game strategy. They organized tasks, responded to threats, and passed useful information among themselves while playing. That makes the run notable as a demonstration of autonomous cooperation within a game environment.
At the same time, the reported result is specific to this Minecraft session. The account describes a successful run by four Gemini 4 Argon agents; it does not establish how they would perform in other games or settings.
The reported cost of the session
The entire session cost $1,620 at the standard Gemini 4 Argon price that takes effect in 2027. That cost accompanied more than a day of play, hundreds of deaths, and thousands of programs written by the agents.
Together, the time, cost, and activity counts give a sense of the resources behind this particular autonomous game run. The central result remains straightforward: four agents coordinated without human input and beat Minecraft.