Four autonomous Gemini agents defeat Minecraft's Ender Dragon in 27 hours

Four agents, one dragon, no humans

Google DeepMind placed four AI agents running its Gemini 4 Argon model into a randomly generated Minecraft survival world and left them to figure it out. Twenty-seven hours and fourteen minutes later, the Ender Dragon — the game's final boss and the usual yardstick for "beating" Minecraft — was dead. According to the post on dev.to describing the experiment, no human intervened at any point along the way.

A deliberately sparse setup

The agents received no assigned roles, no strategy and no purpose-built tooling beyond Mineflayer, a standard JavaScript library for scripting Minecraft bots. Every other action had to come from the agents themselves, in the form of code they wrote and executed on the fly. A fifth agent, called the Observer, stayed outside the game and audited the run while the other four played.

The numbers behind the session

The dev.to post breaks the attempt down into raw activity. Across the run, the four agents wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times. Division of labour appeared on its own — one agent reportedly ended up handling cleanup tasks — and the group coordinated through text, warning one another about threats and recovering equipment from teammates who had been killed.

The session consumed roughly 1.35 billion input tokens, which the post prices at about $1,620 using Argon's standard rates.

Reasoning rather than training

The central claim of the demonstration is that these models were not trained to play Minecraft. There was no task-specific fine-tuning; instead, the agents worked out crafting recipes through reasoning, split mining and building work among themselves, and iterated through a loop of planning, coding, failing and revising. The post argues that this loop is precisely what agentic AI is supposed to deliver in software development and robotics, with Minecraft serving only as the arena.

The limits of the demo

The caveats are stated openly in the post. This was a single run, played on easy difficulty. The model also almost certainly already "knows" Minecraft through its training data, so the agents were reasoning over familiar material rather than discovering the game from zero. And because it has happened only once, there is no evidence yet on how reliably the result could be repeated.

Why it matters

Games have long served as milestones for AI, but earlier headline results — Atari, StarCraft, Go — came from systems trained specifically for their game. This run points at something different: a general-purpose model, given no game-specific training, sustaining coordinated autonomous behaviour for more than a full day.

That sustained behaviour is the hard part of agentic AI. Long-horizon tasks tend to fail not because models lack knowledge but because they drift, lose track of context or cannot recover from their own mistakes. Here, 605 deaths and 2,783 written programs became part of the process rather than the end of it — the agents noticed failures, revised their code and pressed on. The self-organised division of labour, plus an Observer agent auditing the others, also sketches how multi-agent systems might keep themselves in check.

The honest framing is that this is one demonstration, not a benchmark. But as evidence that a general model can hold a plan together across 27 hours of autonomous, code-driven problem-solving — for roughly $1,600 in compute — it is a data point worth watching.