Four Gemini Agents Beat Minecraft Without Anyone Playing: 2,783 Programs, 27 Hours, 605 Deaths
Four AI agents defeated Minecraft’s Ender Dragon with no human input — the first end-to-end autonomous completion of the game, according to a Google DeepMind team’s paper posted to arXiv (Faldor, Singh, and Raileanu) and reported by The Decoder. The four agents ran on Gemini 4 Argon, entered a randomly generated survival world on Easy difficulty with a single paragraph of instructions, and won 27 hours and 14 minutes later after writing 2,783 JavaScript programs, exchanging 229 chat messages, and dying 605 times.
Nobody assigned roles. Nobody wrote game logic. As The Decoder notes, the agents split tasks on their own, warned each other about threats, recovered items from dead teammates, and logged mistakes in a shared memory. A fifth agent — an “Observer” — watched the session as an independent witness for cheating or specification gaming, and found none.
The setup: coding agents, not game policies
This is not a game-playing model. Each of the four players is an instance of the same general-purpose Gemini 4 Argon coding agent, connected to a standard Minecraft server through Mineflayer, an open-source JavaScript library. The agents never see pixels: each one writes a short JavaScript program (“walk to these coordinates and mine every iron ore within 16 blocks”), a bot executes it, and text results come back — position, health, inventory, errors. Then the agent writes the next program. The median program ran for 33.5 seconds; each player had a program running about 40 percent of the time.
The paper’s own accounting of the cost is unusually transparent. The session consumed roughly 400 million input tokens — 91 percent served from context cache — and about 16 million output tokens over roughly 24,000 model calls. At Gemini 4 Argon’s list prices, the four players came to about $520 at the introductory price, $1,040 at the standard price, and the whole session including the Observer and the post-win activity came to about $810 to $1,620.
Why a 2011 video game is standing in for enterprise work
The DeepMind team’s framing is blunt: beating Minecraft is a proxy. Winning requires holding a coherent plan across thousands of dependent steps for more than a day — building tools in the Overworld, running the Nether gauntlet for blaze rods, locating a stronghold, then surviving a dragon fight that alone took nearly 14 hours. That is the same shape as a long-horizon engineering project: dependencies, failures, recovery, coordination.
The honest caveats matter, and the paper lists them. Mineflayer feeds the agents structured state — exact positions, inventories, even the dragon’s internal state — that a human player would have to see or infer, so the perception problem is sidestepped rather than solved. Easy difficulty softens damage. Gemini 4 Argon knows Minecraft from its training data; the test is whether a general agent can convert existing knowledge into working programs, not whether it can handle a novel game. And it is a single run — the first launch they tried — so nobody knows the success rate. Twenty-two percent of programs still ended in an error, and the game server restarted six times mid-run.
From game benchmarks to the pattern that matters
The prior benchmarks against which this lands are worth the comparison. Video PreTraining crafted diamond tools in 2022. DreamerV3 collected diamonds with reinforcement learning in 2025. MineDojo listed dragon-slaying as one of its “moonshot” goals, with purpose-built architectures behind each. The Gemini 4 Argon run beat the entire game with a general model that was never trained on Minecraft specifically — using only the standard bot library and the same tools it does software work with.
That inversion — general agents swallowing purpose-built ones — has been the story of agentic AI this year, and it matches the pattern in earlier coverage of agents failing open-ended research while excelling bounded engineering work. A multi-day, self-directed engineering project with teammates is the enterprise dream scenario; the Minecraft run shows the coordination primitives work, while earlier benchmark work showed the same agents stalling when goals get vague. Minecraft sits at the tractable end precisely because the goal is crisp, the environment is checkable, and death is cheap. The gap between “beat a game in 27 hours” and “run a business process unsupervised” is where the next few years of agent work will live.
For the AI-in-games thread on our earlier coverage of game-playing models, Minecraft joins Pokémon and Civilization VI as the standard proving grounds — but this run is the first where the agents were also the engineers.
Every program, message, and game event from the run has been released: the team published a dataset on Hugging Face alongside the paper, licensed CC BY 4.0.
