How Google DeepMind's Gemini Agents Beat Minecraft: The System Behind It

Google DeepMind researchers ran four Gemini 4 Argon agents in a randomly generated Minecraft survival world, on Easy, with one goal: defeat the Ender Dragon. Each agent controlled a player by writing JavaScript through the open-source Mineflayer library, coordinating through chat and shared notes. They won in 27 hours 14 minutes, with no assigned roles or human help.

Who built it and what they claim

The work is described in an arXiv preprint titled "A team of Gemini 4 Argon agents beats Minecraft", submitted October 15, 2026 by Google DeepMind researchers Maxence Faldor, Satinder Singh, and Roberta Raileanu. The Next Web and others covered it from October 20 onward. The authors write that, to their knowledge, it is "the first end-to-end autonomous completion of Minecraft by AI agents" using only the standard Mineflayer library. Their stated reason for caring is not games: they argue it shows agents can keep a coherent plan across thousands of dependent steps, which matters for software engineering and enterprise work. The preprint has not been peer reviewed.

The architecture, piece by piece

The interesting part for builders is how little game-specific scaffolding there was.

Action space: code, not keystrokes

Each agent controlled one player through Mineflayer, an open-source JavaScript library for Minecraft bots. The agents did not look at the screen or press keys. They wrote short programs (walk here, mine nearby iron, craft this) and, after each one, read back position, health, inventory, or the error. The Next Web reports the programs ran in a sandbox where each one could control only its own player.

Team: four peers, no roles

The researchers assigned no roles or strategy. Instructions covered verifying actions, building reusable tools, keeping lessons in shared memory, and using team chat. Roles emerged without being assigned: one scouted food, beds and flint, one smelted iron and tools and found the fortress and stronghold, one prepared the portal, and one gathered diamonds and landed the final blow.

Coordination: chat plus shared notes

The agents exchanged 229 chat messages and kept shared notes of coordinates, warnings, and tool lessons. Shared chests pooled resources. Teammates recovered each other's items after deaths.

Oversight: an Observer agent

A fifth agent, the Observer, did not play. It checked for cheating against recorded events and published live and summary reports. It found none, per The Next Web. Using a separate, non-acting agent as an auditor is a pattern worth copying in any multi-agent system.

What it cost, and where the tokens went

The AI Insider reports about 24,000 model calls, about 1.35 billion input tokens, and about 16 million output tokens through the win. The researchers estimated about $520 at Gemini 4 Argon's introductory list price, or about $1,040 at standard price, and roughly $810 to $1,620 for the whole session including the Observer.

Do the arithmetic on those reported figures and the shape is clear: roughly 84 input tokens for every output token, or around 56,000 input tokens per call on average. The agents spent most of their budget re-reading context (notes, history, game state), not writing code. That is typical of long-running agents, and it is why the cost of a long run is governed by what gets re-sent on every call, not by how much the model writes. Retrieval of only the relevant notes, and caching of stable prefixes, are the standard levers.

What the result does not prove

The coverage, and the authors, list real limits.

  • One run. The researchers describe a single successful run, their first launch. Reliability across worlds and seeds is unknown.
  • Easy difficulty. Normal or Hard would change the combat budget substantially.
  • Superhuman sensors. Mineflayer reports exact nearby blocks and creatures, including through solid rock, and the dragon's internal state. A pathfinding plugin handled routing. Screen-and-keyboard play is acknowledged as much harder.
  • No single-agent baseline. Without one, how much the team structure contributed is unmeasured.
  • Prior knowledge. The model already knew Minecraft and Mineflayer from training.
  • Not the first dragon kill. Earlier community projects reported dragon defeats; the claimed distinction is the absence of game-specific tools and the end-to-end setup.

Read it as evidence of sustained, self-organised coordination over a day, through a text and code interface. Not as evidence that agents play like people.

Lessons for building multi-agent systems

Three design choices carry over directly. Give agents a code-based action space with structured feedback rather than pixel control where an API exists. Make memory shared and written down, so lessons survive individual agents. And put an independent auditor beside the actors. Faldor said on X that DeepMind plans to apply the same setup to coding, science, and robotics, which is the real test.

Frequently asked questions

Which model did the DeepMind Minecraft agents use?
All four playing agents, and the separate Observer, ran on Gemini 4 Argon, according to the arXiv preprint and coverage in The Next Web. Each agent controlled its own player by writing JavaScript programs through the open-source Mineflayer library, with no screen input or keyboard control, and coordinated with teammates through in-game chat and shared notes.
How long did it take the AI agents to beat Minecraft?
The team defeated the Ender Dragon 27 hours and 14 minutes after the first program ran, starting from a randomly generated survival world on Easy difficulty. Along the way the agents wrote 2,783 programs and died 605 times, 69 of those deaths to the dragon, per the preprint as reported by The Next Web and The AI Insider.
Did humans help the agents?
The researchers say no. They supplied no roles, strategy, or game-specific tools, and watched a live dashboard without intervening, according to The AI Insider. The instructions only covered general practices such as verifying actions, building reusable tools, and keeping shared notes. A separate Observer agent checked for cheating and reported finding none.
How much did the Minecraft run cost?
The researchers estimated about $520 in model fees through the win at Gemini 4 Argon's introductory list price, or about $1,040 at standard price. Including the Observer agent for the whole session, estimates were about $810 or $1,620. These are list-price estimates published by the researchers, not invoices, so treat them as approximate.
Is this the first time AI beat the Ender Dragon?
The authors claim the first end-to-end autonomous completion using only the standard Mineflayer library. Coverage notes that earlier community projects reported dragon kills using game-specific tools or fixed, easier setups. The distinction is the conditions, not the kill itself, and the preprint is not yet peer reviewed.