Google DeepMind beats Minecraft with four Gemini 4 Argon agents

No one taught them how to play, and no one assigned the work in advance. Four AIs wrote 2,783 programs themselves, spent 27 hours and 14 minutes, and ultimately defeated the Ender Dragon.

On October 15, 2026, researchers at Google DeepMind published a paper, "A team of Gemini 4 Argon agents beats Minecraft."

The researchers did something quite crazy: they placed four general-purpose AI coding agents trained by Gemini 4 Argon into a randomly generated Minecraft survival world.

There was no walkthrough route, no one had pre-assigned who would handle mining and who would handle fighting mobs, and there was no human-written program for clearing the game.

For veteran players of Minecraft (MC), defeating the Ender Dragon is nothing new.

Players must go through an entire chain of interdependent tasks.

Starting from a world with nothing, they chop trees, mine, craft weapons and equipment, then enter the perilous Nether, collect blaze rods and ender pearls, find the underground End fortress, and finally arrive in the End to destroy the ten end crystals that heal the Ender Dragon.

The researchers used Minecraft Java Edition 1.20.4, enabled survival mode, selected Easy difficulty, generated the map randomly, and did not enable cheat commands.

The four Gemini 4 Argon agents each controlled one character, using the open-source tool Mineflayer to obtain that character's position, inventory contents, health, and surroundings, and wrote JavaScript code themselves to direct the character's behavior.

Although building a similar basic environment is not difficult, fully reproducing it would still require the corresponding model, execution framework, and memory system; one cannot simply download some plugins and make an AI go fight the dragon on its own.

The researchers did not prepare a game walkthrough for them. About ten minutes into the game, janesick proposed in the chat channel that they divide the work. Within a short time, the four AIs gradually formed their own responsibilities: gyldenia handled resource logistics, janesick handled tool crafting and exploration, auratian handled key technical challenges, and bulicame handled resource prospecting and the final battle.

The biggest difference from many previous agent demonstrations is that the role division was not written into the prompt from the start, but gradually formed through cooperation. Moreover, these responsibilities were not fixed. Depending on the game's progress, the four AIs would also change their tasks accordingly.

Less than two hours after the game started, the AI team had obtained iron gear and diamonds and built a portal to the Nether.

When playing games, a character's death is usually a failure.

But in this experiment, the four AIs died a cumulative 605 times, which precisely demonstrated an important capability of long-term autonomous agents: even if a single action fails, the entire task does not have to fail with it.

When JaneSick died carrying 12 blaze powders and the dropped items were destroyed by a ghast fireball, what was lost was not just one character's life, but resources painstakingly collected over the previous few hours.

auratian reminded teammates about the danger of transporting supplies, and janesick summarized the cause of the accident and saved the lesson to shared memory.

Next, the cause of the failure had to be analyzed to see whether it was due to a code error, a pathing error, insufficient resources, or an incorrect strategy established earlier.

This made a character death no longer just a loss in the game, but an experience that the whole team could reuse, just as when you lose enough in a game you always accumulate some experience to use next time.

Experience can serve not only one agent but also be used by other agents, and it works during a task that lasts for several days, enabling the AI to continue advancing its long-term goals even after many local failures.

Entering the game's "infinite flow"

Over the past few years, the report cards of large models have mostly come from answering questions, writing code, taking exams, and setting new scores on various benchmarks. Entering game environments places more compound demands on large models' capabilities.

The four agents could maintain consistent goals in the same complex environment, and after tens of thousands of program calls and a large number of failures, they finally completed their work.

What it tests is a complete, continuous execution process, not one-shot reasoning ability.

The four Gemini agents used Mineflayer to read the game environment and converted tasks into JavaScript programs; after execution, they judged whether the task was complete based on the results. When a character encountered an obstacle, the program would produce an error, and at that point the cause had to be analyzed and the code rewritten.

The original paper's data shows that a program ran for about 33.5 seconds. Throughout the game, each agent spent about 40% of its time writing game programs, and the remaining 60% analyzing results, reasoning, and writing the next set of code.

Even if the model has a very large context window, it cannot be guaranteed to remember everything from the past.

Taking Gemini 4 Argon as an example, according to Google's official developer documentation, this model can reach a context length of about one million tokens and has longer output, tool-calling, and reasoning capabilities.

From the beginning until the Ender Dragon was defeated, the four players generated a total of about 24,000 model calls, processed about 1.35 billion input tokens, and produced about 16 million output tokens, of which 91% of the input tokens came from context caching.

Such a large interaction history cannot fit all the information into every model call.

Back in 2023, Voyager's research had already used Minecraft to test this idea: using an automatic curriculum, a code skill library, and execution feedback to enable a GPT-4 agent to continuously explore and accumulate skills.

In this experiment, the researchers' approach was to equip each agent with an independent and persistent memory, and to open up team-shared memory. As the context continued to grow, the agents would perform context compaction, recording the current task, important progress, unfinished plans, and important lessons into external memory.

How far are we still from "gaming freedom"?

Speaking of it, AI autonomously beating Minecraft is not some novelty that only appeared in 2026.

As early as May 24, 2021, the open-source robot project AltoClef had already achieved fully automatic completion of Minecraft. However, it used a client mod and Baritone automatic pathfinding, relying on task logic pre-written by the developers to complete the game.

After that, game AI research developed in a more general direction, and many new models emerged, such as Voyager and others.

The Gemini 4 Argon experiment is not about teaching AI to play games, but about using programs to implement various functions on the basis of already having general programming capabilities.

The first result reported by DeepMind was that, under the standard Mineflayer interface, without any preview map, and without any tools or task strategies designed specifically for the game, ordinary agents completed the game from start to finish on their own.

In June 2026, KRAFTON and NVIDIA launched a limited-time test event for PUBG Ally, adding AI as teammates to help players complete game tasks. Players could use voice commands to have the AI search for weapons, collect supplies, and participate in combat.

This likewise used the understanding and planning capabilities of large models to connect the actual operation interfaces of the game system.

Judging from the results, AI cannot yet be considered stable. About 22% of the game programs encountered errors, and the server was restarted 6 times.

They ultimately succeeded, but this was only one successful experiment. The paper did not prove that they can reliably complete the game through enough repeated tests on random maps.

In addition, with about 24,000 model calls within 27 hours, calculating at discounted API prices, the completion cost would still be about $520. If one wanted to use it for boosting, then this account would first need to be calculated.

Players who cannot wait to use large models to play games in their place can hold off for now; at least for the time being, game controllers and keyboards remain firmly in human hands.

The paper link is here, so everyone can take it and try to replicate it 👇

This article comes from the WeChat public account "APPSO", author: APPSO, which discovers tomorrow's products, and is published by 36Kr with authorization.