Google researchers report four AI agents beat Minecraft

The researchers’ paper describes four Gemini 4 Argon agents in a Minecraft survival world, each controlling a player through the open-source Mineflayer library.[3] Their assigned objective was to defeat the Ender Dragon.[3] The agents coordinated using in-game chat and shared memory, divided work among themselves, passed on what they learned and recovered one another’s items.[3] The paper was submitted to arXiv on October 15, 2026.[2] IGN reports that the agents exchanged 229 messages during the playthrough.[1]

The researchers say they did not assign roles or write the agents’ game logic; instead, the agents planned, wrote and debugged their own programs.[3] Researchers could monitor the run and pause or stop it, but did not intervene.[3] The run took place in a randomly generated world on Easy difficulty, with cheat commands disabled and no human input.[3] The researchers also say they used the standard Mineflayer library, without pre-scouted seeds or human-written, game-specific tools.[3] These details describe the conditions of this experiment, not a test across every Minecraft setting.

The agents defeated the Ender Dragon after 27 hours and 14 minutes.[3] IGN reports that reaching The End took around 13.5 hours.[1] The agents wrote 2,783 programs and died 605 times before completing the task.[3] Those figures make the result relevant to teams assessing autonomous systems: completion alone does not describe the time, iteration or failures involved. The evidence here documents one run, so use its duration and setbacks as details of this experiment rather than as a general performance benchmark.

IGN reports that the agents used 1.35 billion input tokens and says that usage would cost a normal user around $1,620.[1] Tom’s Hardware gives a different estimate, putting the token use at about $520.[4] The evidence provides both figures but does not explain the difference, so they should not be treated as a single agreed cost. The researchers say the experiment was intended to study multi-agent coordination in a setting where each step is visible.[1] They describe the result, to their knowledge, as the first end-to-end autonomous Minecraft completion by AI agents.[3]