Google DeepMind researchers have shown that four Gemini 4 Argon AI agents can autonomously complete Minecraft, defeating the game's final boss after more than 27 hours of continuous collaboration.

According to a research paper released Thursday, the agents wrote and executed 2,783 JavaScript programs, exchanged 229 messages and died 605 times before taking down the Ender Dragon.

The run took place in a randomly generated Minecraft survival world on Easy difficulty, a setting that reduces the damage the game's hostile creatures deal. Researchers gave the agents a single objective and did not assign roles, develop gaming strategies or intervene at any point during the attempt. DeepMind described the outcome as the first known start-to-finish autonomous completion of Minecraft using general purpose AI agents, without custom gaming tools or predetermined world configurations.

Agents coordinated and built their own strategies

Each agent controlled a Minecraft character through Mineflayer, an open source JavaScript library that lets programs interact with the game. Rather than pressing keys or interpreting gameplay visually, the agents wrote code to navigate the world, gather resources, craft equipment and fight enemies.

A division of labor emerged within the first ten minutes. One agent gathered supplies, another focused on tools and exploration, and the others prepared for the game's more hazardous environments. The agents also maintained shared memory, recording locations, strategies and lessons from earlier failures. Over the course of the run, they developed 42 reusable helper functions to automate movement, resource collection and survival.

One agent calculated the location of a hidden stronghold by tracking the movement of two Eyes of Ender — in-game items whose flight paths point toward the stronghold — and determining where their trajectories intersected, an estimate that landed within 43 meters of the actual portal room.

The team ran into significant difficulties after entering the End, where the dragon repeatedly killed members. The breakthrough came when the agents identified the ten crystals that continuously restored the dragon's health. They devised a strategy using stacked dirt blocks to reach and destroy the crystals before finally defeating the boss.

The full run took 27 hours and 14 minutes, including six server restarts.

A test of long-horizon autonomous coordination

The experiment arrives amid a broader push across the AI industry toward agents — systems that plan and execute multi-step tasks with limited human supervision — where sustaining reliable performance over long stretches of dependent work remains a central challenge.

Beyond completing the game, the experiment tested whether general purpose AI systems could coordinate across thousands of dependent tasks without continuous human guidance. The four agents collectively generated roughly 63,313 lines of JavaScript and held a shared strategy over more than a day of activity.

According to the researchers, the agents processed approximately 1.35 billion input tokens — the units of text and code a model reads — and generated 16 million output tokens across roughly 24,000 model calls. At Gemini 4 Argon's introductory API pricing, completing the game cost about $520; at standard pricing, the figure rises to $1,040. Because models are billed by token volume, costs scale directly with the length and complexity of an agent run.

The researchers argued that the result demonstrates capabilities relevant to complex software engineering and enterprise workflows, where AI systems must plan, coordinate and recover from failures over extended periods. The run was nonetheless conducted in a controlled gaming environment, with programmatic access to game information rather than visual interaction. The researchers also acknowledged that the experiment does not establish whether such capabilities would transfer directly to real-world environments, where tasks may be less predictable and mistakes more costly.

That caveat frames what to watch next: whether follow-up work can show this kind of extended, unsupervised coordination holding up outside a game engine, where actions are harder to reverse and errors carry real consequences.

Even, the findings point to progress in AI systems capable of holding objectives, building tools and collaborating autonomously across lengthy sequences of actions.