Four AI agents running on Google's Gemini 4 Argon have beaten Minecraft without any human input, slaying the Ender Dragon in a fully autonomous run that lasted 27 hours and 14 minutes. The Google DeepMind team dropped the agents into a randomly generated world without assigning roles, strategies or game-specific tools, according to a paper by Faldor et al. covered by The Decoder — and the agents organized themselves, cooperated, and won.

Minecraft has long been a stubborn benchmark for AI. Unlike chess or Go, it demands long-horizon planning, open-ended exploration, tool crafting, and coordination — all in a world that changes every time you spawn into it. Beating the game's final boss has historically required either human guidance or heavily scripted training setups. This run, by contrast, began with four agents and nothing else.

The Run in Numbers

The statistics from the session read like the log of a chaotic but ultimately successful expedition:

  • 27 hours, 14 minutes of continuous autonomous play
  • 2,783 JavaScript programs written by the agents themselves
  • 229 chat messages exchanged between the four agents
  • 605 deaths along the way

That last figure deserves emphasis. The agents died more than six hundred times before completing the game, and per Tom's Hardware's coverage of the paper, they did it on the game's Easy difficulty setting. The same report estimated the token spend for the run at roughly $520, while The Decoder, calculating at Gemini 4 Argon's standard pricing that takes effect in 2027, put the total at $1,620. Either way, the cost of an end-to-end Minecraft completion by general-purpose AI agents is now measured in hundreds of dollars — not the millions invested in the purpose-built game AIs of previous years.

Emergent Teamwork, No Scripts

What makes the result notable is not just that the agents won, but how. With no roles assigned, the four agents split tasks among themselves, The Decoder reports. They warned each other about threats, recovered items from dead teammates, and logged mistakes in a shared memory that let the group avoid repeating errors.

In other words, the coordination behavior that game developers usually hard-code into NPC teams emerged on its own from four instances of the same model, a shared world, and a communication channel. The 2,783 self-written JavaScript programs function as an ever-growing skill library: once an agent figures out how to smelt iron or defeat a miniboss, that knowledge persists as reusable code rather than living only in the model's context window.

The shared-memory detail matters for anyone following agent research. Teams of agents that can accumulate and transmit lessons — rather than restarting from zero after every failure — address one of the central weaknesses of today's AI agents, which tend to lose hard-won progress the moment a session ends.

From Purpose-Built Champions to General Agents

The result is the latest example of a pattern that has been accelerating across the field. A decade ago, beating world-class video game AI required bespoke systems trained from scratch for one game: DeepMind's AlphaGo for Go, OpenAI Five for Dota 2, AlphaStar for StarCraft II. Each was a monumental engineering effort that could do exactly one thing.

General-purpose agents are now clearing game challenges that once demanded years of specialized training — and they do it with the same model that can answer emails or write code. The Decoder notes that Minecraft is just the latest example of agentic AI models beating video games that used to require purpose-built systems. Each such result shortens the distance between "impressive demo" and "generally capable assistant that can pursue a long goal in a messy environment."

That is precisely why Minecraft has been such a persistent research testbed: surviving creepers is incidental, but sustaining goal-directed behavior across tens of hours in an open world is a decent proxy for the kind of autonomy AI labs are trying to build.

Caveats and Context

There are caveats worth keeping in mind. The run used Easy difficulty, per Tom's Hardware, and Minecraft — while famously open-ended — remains a game with well-defined rules, deterministic physics and generous respawn mechanics. The 605 deaths would be fatal in most real-world settings; here they cost time and tokens.

The compute and cost figures are also worth reading carefully: the two published estimates differ by a factor of three depending on the pricing model used, a reminder that "how much did the agent run cost" has no single answer until inference prices stabilize.

Still, none of this diminishes the core result. Four general-purpose agents, given nothing but a world and each other, self-organized into an effective team and finished one of gaming's longest-running AI challenges in barely more than a day.

What It Signals for Agentic AI

For DeepMind, the demonstration doubles as a showcase for Gemini 4 Argon's agentic capabilities — the model family Google is positioning for exactly this kind of long-horizon, tool-using work. As enterprises move AI agents from answering questions to executing multi-day workflows, benchmarks like this one offer a glimpse of what current models can sustain when the objective is long, the environment is adversarial, and failure is measured in hundreds of small deaths rather than one big one.

The Ender Dragon has fallen to machines before, with heavy human scaffolding. This time, four agents walked into a random world alone — and walked out as champions.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →