DeepMind paper: Gemini 4 Argon agents autonomously beat Minecraft, researchers say

A new Google DeepMind preprint says four Gemini 4 Argon agents autonomously completed Minecraft by defeating the Ender Dragon after 27 hours and 14 minutes of planning, coding and coordination, a notable result for long-horizon AI agents that still comes with important caveats.

In the paper, posted Thursday on arXiv and not peer reviewed, Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu wrote: “We placed four Gemini 4 Argon agents in a sandboxed Minecraft survival world and gave them one goal: defeat the Ender Dragon.” According to the paper, the agents reached that goal after writing and running 2,783 programs before the win, dying 605 times and exchanging 229 in-game chat messages. Researchers say no human intervened during the run, though they could watch on a live dashboard and pause or stop it.

Minecraft matters to AI researchers because finishing the game is a long chain of dependent tasks, not a single boss fight. Players must gather resources, craft tools, enter the Nether, find a fortress, make Eyes of Ender, locate a stronghold, reach The End and then survive the final battle. That makes it a useful test of whether an AI system can plan over many hours, recover from mistakes and coordinate multiple steps.

The setup described in the preprint used four Gemini 4 Argon agents in a randomly generated Minecraft Java Edition 1.20.4 survival world on Easy difficulty, with cheat commands disabled. The agents controlled players through Mineflayer, an open-source Node.js library for programmatic Minecraft control. They coordinated through in-game chat and a shared persistent memory system.

The paper says no human wrote game logic or assigned roles. Instead, the agents planned, wrote and debugged their own JavaScript programs during the session. In total, they produced 2,839 programs, including work after the win, and 63,313 distinct lines of JavaScript. The researchers also say the agents’ programs could not make network requests, start processes or write files, limiting them to the tools available inside the sandboxed setup.

Those details matter because this was not human-like play. The agents were not using screen pixels, a keyboard or a mouse. They used Mineflayer’s structured access to game state, which the paper acknowledges exposes privileged internal information. The setup also used a pathfinding plugin. That makes the result different from embodied AI benchmarks or from a person playing the game directly.

The authors nevertheless argue the result is a milestone in general-purpose coding agents. “To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans,” they wrote.

That “first” claim needs qualification. Autonomous Minecraft completions have been reported before, most notably by the Altoclef project. Its GitHub README says it “Became the first bot to beat Minecraft fully autonomously on May 24, 2021.” The distinction is in how the systems were built: Altoclef used a client-side mod, Baritone pathfinding and hand-authored task logic, while DeepMind’s preprint describes large language model agents that wrote code at runtime and coordinated through more general tools.

So the clearer claim is not that no bot has ever finished Minecraft before, but that DeepMind is presenting a different kind of system: a team of general-purpose coding agents using standard Mineflayer rather than bespoke, human-authored game logic.

The paper, titled “A team of Gemini 4 Argon agents beats Minecraft,” carries arXiv ID 2610.22319 and was submitted Oct. 15, 2026. The result also serves as an early showcase for Gemini 4 Argon, which Google introduced Sept. 30 as a reasoning model aimed at long-horizon coding and multistep autonomous tasks. As Google put it at the time, “Gemini 4 Argon is our flagship reasoning model.”