Four Google Gemini 4 Argon AI agents autonomously defeated the Ender Dragon in Minecraft after 27 hours, coordinating via chat and code without human intervention.
By Hermes-Vector AI Desk
Google researchers have demonstrated that four Gemini 4 Argon AI agents can autonomously complete Minecraft, defeating the game’s final boss, the Ender Dragon, without human help. The achievement, detailed in a research paper, marks what the team describes as the first end-to-end autonomous completion of the game by AI agents using only standard libraries.
The experiment involved placing four agents into a randomly generated Minecraft Java Edition 1.20.4 survival world on Easy difficulty. The agents were given a single objective: defeat the Ender Dragon. No specific roles or game strategies were assigned beforehand.
According to the researchers, the agents coordinated through in-game chat and shared memory. "Each agent controlled a player via an open-source library called Mineflayer and coordinated with its teammates through in-game chat and shared memory," the researchers said. "No one assigned roles or wrote game logic. The agents planned, wrote, and debugged every program themselves."
The agents quickly divided tasks. One agent focused on tool creation, another on gathering food, while others handled exploration and shelter construction. They shared combat tips and assisted each other in recovering lost items after deaths. A fifth agent, referred to as the Observer, monitored the session to ensure no cheating occurred but did not intervene in gameplay.
The entire run took 27 hours and 14 minutes. During this time, the agents wrote and executed 2,783 JavaScript programs and exchanged 229 chat messages. The process was not without failure; the agents died 605 times, and the game server restarted six times, pausing the world for approximately 20 minutes in total.
To reach the End dimension, the agents had to gather Ender Pearls and Ender Rods, craft Eyes of Ender, and locate a Stronghold. One agent estimated the Stronghold’s position to within 43 meters of the actual portal room by tracking the movement of two Eyes of Ender. Once in the End, the agents identified 10 crystals that restored the dragon’s health and developed a strategy using stacked dirt blocks to destroy them.
The agents generated approximately 63,313 distinct lines of JavaScript and developed 42 helper functions. About 22% of the programs ended in an error, highlighting the significant amount of recovery and debugging required during the run.
The experiment processed approximately 1.35 billion input tokens and generated 16 million output tokens across roughly 24,000 model calls. About 91% of input tokens were served from cache. Applying API list prices, the cost was estimated at about $520 at introductory rates or $1,040 at standard rates. IGN reported a different cost estimate, stating the tokens would cost a normal user around $1,620, but the research paper and other outlets cite the $520 to $1,040 range based on specific API pricing tiers.
The agents did not interpret screenshots visually. Instead, they received structured information about their characters and the game world, including data on blocks behind solid rock and the dragon’s internal state. A pathfinding plugin was used to plan routes.
The researchers stated that the goal was to study how teams of agents coordinate in a visible setting to understand such systems before they are trusted with real-world tasks. "The skills required to solve Minecraft such as coordinating a team, recovering from failure, and pursuing a goal for more than a day, are also what agents need to work reliably alongside people," they said.
However, the experiment was a single attempt on Easy difficulty with programmatic access to game data. The researchers acknowledged that the results do not establish how reliably the agents could repeat the run or whether these capabilities would transfer directly to unpredictable real-world environments. The current human speedrun record for Minecraft is six minutes and 36 seconds, indicating that while the AI demonstrated autonomy, it is not yet efficient compared to human players.
Sources
This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.
The day's checked news, by email
One short email a day with the top stories. Free, and you can unsubscribe in one click.
Or follow for the latest news:
No comments yet. Start the discussion.
Microsoft introduced new security features to monitor AI agents, warning that autonomous systems can pose insider risks even without being hacked. The company urges separating intelligence supply from control authority.
The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek AI products from IT systems and barred future procurements, citing cybersecurity risks.