Four Gemini 4 Argon agents beat Minecraft with no human help

Four AI agents walk into a freshly generated Minecraft world with empty hands and one sentence of instruction: defeat the Ender Dragon. Nobody tells them who should mine, who should fight or where the stronghold is. Twenty seven hours and fourteen minutes later, the dragon is dead.

That is the result of a new paper titled "A team of Gemini 4 Argon agents beats Minecraft", posted to arXiv on 15 October 2026 by Maxence Faldor, Satinder Singh and Roberta Raileanu. According to The Decoder, the work comes from a Google DeepMind team. The four players were all instances of the same general purpose coding agent running on Gemini 4 Argon, Google's newest frontier model, which became generally available on 13 October.

The authors call it, to their knowledge, the first end to end autonomous completion of Minecraft by AI agents that used only the standard open source Mineflayer library, with no pre scouted world seeds and no game specific tools written by humans. Along the way the team wrote 2,783 programs, exchanged 229 chat messages and died 605 times.

Why beating Minecraft is hard for an AI

Minecraft looks like a toy, but finishing it from a random start is a long chain of dependent steps across three worlds. Players have to gather wood and stone, craft tools, find iron and diamonds, and build a portal into the Nether. There they must find a fortress, kill blazes for blaze rods and collect ender pearls from Endermen. Back in the Overworld, they combine both into Eyes of Ender, throw them to locate an underground stronghold, and use twelve eyes to light the portal to the End. Only then does the actual boss fight start, against a dragon that keeps healing itself from crystals on top of tall obsidian pillars.

Researchers have treated this as a moonshot for years. The MineDojo project in 2022 listed defeating the Ender Dragon from a bare handed start in a fresh world as one of its hardest goals. Earlier systems such as OpenAI's Video PreTraining or DreamerV3 got as far as diamond tools, and they needed huge amounts of gameplay data or training experience. Language model agents like Voyager reached similar points with a growing library of skills and a curriculum designed by humans.

What changes here is not the game. It is the kind of agent. Nothing in this setup was built for Minecraft.

How the agents played

Each of the four players was driven by Google's general purpose Gemini 4 Argon coding agent, the same agent the team uses for software work. It ran on its own cloud machine with a code workspace, version control, a shell, file editing, web search and a persistent memory with a private part and a part shared by the team.

The agents did not see pixels or press keys. Instead, they controlled their characters through Mineflayer, an open source JavaScript library for Minecraft bots, plus its pathfinding plugin. An agent acted by writing a short program, for example "walk to these coordinates and mine every iron ore within 16 blocks", sending it to its bot, and reading the text that came back: position, health, inventory, what it found or the error it hit. Then it wrote the next program.

The median program ran for 33.5 seconds. Each player had a program running about 40 percent of the time. The rest went into reading results, thinking and writing code.

The instructions about the game were one paragraph. The objective was to defeat the Ender Dragon, starting from spawn in a shared survival world, by any means allowed in survival mode, and "the players decide how to share the work." Everything else came from general rules the agent follows in any environment: check the result of each action, build reusable tools, write lessons into shared memory and report honestly.

The world ran on Minecraft Java Edition 1.20.4 in survival mode on Easy difficulty, with cheat commands disabled and no human input during the run. A fifth agent, the Observer, watched as an independent witness. It never played. Its job was to check the players' claims against the logs and to flag cheating, exploited bugs or invented results. It flagged none.

The 27 hour run, step by step

The paper reads a bit like a match report, and the timeline shows how messy and human the process looked.

Hour 0: a team forms. Ten minutes after the first program, the agent called janesick proposed a division of labour in the chat. Within a minute, auratian answered with a different plan, and gyldenia replied "Copy that!". The roles settled fast: gyldenia gathered food, beds and flint, bulicame went down for diamonds, janesick smelted iron and made tools, and auratian prepared the Nether portal. After 1 hour and 42 minutes the portal was lit.

Hours 2 to 11: the Nether. janesick found a fortress with a blaze spawner. The other three hunted Endermen in a warped forest using a classic trick: they built shelters only two blocks high. An Enderman that is looked at teleports toward the player, but at almost three blocks tall it cannot fit under the low ceiling. About five and a half hours in, the team had 15 ender pearls in a shared chest. Getting the blaze powder home took more than six hours. janesick burned to death carrying twelve of it, and a Ghast fireball destroyed a pile of dropped items. After that, auratian posted a new rule for everyone, and janesick added the root cause and a fix two minutes later.

Hours 10 to 13: finding the stronghold. janesick threw two Eyes of Ender from different spots, sampled each eye's flight every 50 milliseconds and calculated where the two lines crossed: 1.3 kilometres from the village. The estimate landed 43 metres from the actual portal room. Then auratian drowned carrying all twelve eyes. Two teammates sent bulicame the exact spot and the minute the items would disappear, and bulicame picked up all twelve with two and a half minutes to spare. At 13 hours 17 minutes, the End portal was lit.

Hours 13 to 27: learning the dragon. This was the hard part. The dragon killed the team 69 times. Between attempts, the agents read the dragon's internal state from the game and wrote down rules: never wait on the arrival platform, approach only while it is perched, and leave the End to reset it when it gets stuck. To leave, they jumped into the void on purpose, which sends a player back to its bed, a shortcut human players use too.

The breakthrough came about 21 hours in. auratian pointed out that the ten crystals on the pillars kept healing the dragon to full health, and worked out how to reach them without a pickaxe: stack dirt to climb beside each pillar, punch the crystal, dig back down. Over four hours and forty minutes the team destroyed all ten. Thirty nine minutes after the last one fell, the dragon died, with bulicame down to 2.5 of its 20 health points.

Teamwork nobody programmed

The most interesting part of the paper is not the win itself but how the agents organised the work. Their instructions never mentioned roles, chests or rescues. Over 229 messages, the four converged on habits familiar from good human teams:

• Specialisation. Each agent settled into a role. gyldenia gathered supplies and later destroyed six of the ten crystals, janesick built tools and found both the fortress and the stronghold, auratian lit the portal and solved the crystal problem, and bulicame landed the final blow.
• Shared memory. The agents posted coordinates of chests, portals and dangers. Nineteen minutes in, gyldenia warned the others about a crafting quirk in Mineflayer. A helper function with the fix ended up in 655 programs written by all four agents.
• Their own tools. The team kept 146 pieces of state and code alive between programs, including 42 helper functions: a background loop that keeps a bot afloat in water, timed walking, item counting and a counter that stops leftover code from a timed out program. That last one appears in 1,519 programs.
• Looking out for each other. Urgent warnings about threats and recovering a teammate's items after a death were routine.
• Self correction. When auratian's program announced a success that had not happened, auratian corrected it two minutes later, calling the message "an unguarded script branch". When janesick reported coming back with twelve blaze powder one second after burning to death, gyldenia flagged the contradiction within three minutes.

That last point matters beyond games. Agents that confidently report results they did not achieve are one of the most common failures in real deployments. Here, the team caught its own false claims without any human in the loop.

What it cost

The paper gives a rough bill. Up to the win, the four players processed about 1.35 billion input tokens, 91 percent of them served from the context cache, and generated about 16 million output tokens, including 4 million thinking tokens, over roughly 24,000 model calls.

At Gemini 4 Argon's introductory public API price, that comes to about 520 dollars, or 1,040 dollars at the standard price. The whole session, including the Observer and a day of activity after the win, came to about 810 dollars, or 1,620 dollars at the standard price. The authors stress that these are estimates from list prices and token counters, not what it costs Google to serve the model.

For context, the HowLongToBeat median for Minecraft's main story is 45 hours for human players, and the random seed speedrun record is under seven minutes. The agents were slower than a speedrunner and faster than a typical player, but they also played as a team of four, through code instead of a screen.

The limits the authors admit

The paper is unusually upfront about what the result does not show, and these caveats are worth reading before anyone calls Minecraft "solved".

• Code, not eyes. Mineflayer tells the agents exactly which blocks and creatures are around them, including blocks hidden behind rock and the dragon's internal state. A human has to see and interpret all of that, or cannot see it at all. The pathfinding plugin also plans routes around obstacles. Playing from pixels with a keyboard and mouse remains much harder.
• Easy mode. Monsters deal less damage than on Normal or Hard, and hunger cannot kill a player.
• Prior knowledge. Gemini 4 Argon knows a lot about Minecraft and Mineflayer from its training data. The test shows that a coding agent can turn that knowledge into working programs and teamwork over a day, not that it can learn an unknown game.
• Lots of failure. 605 deaths, 22 percent of programs ending in an error, and almost 14 hours spent in the End.
• One run. It was the first launch the team tried, which says nothing yet about how reliably the setup succeeds.

The authors also note two recent community projects that reported dragon kills: a single Claude Opus 5 agent using 37 high level Minecraft tools on Normal difficulty, and a planner paired with an action model on a fixed speedrun seed in Peaceful mode. The difference here is how little the agents were given: a standard library, one paragraph of objective and their general working rules.

Why this matters outside of games

The point is not really Minecraft, and the paper says so directly: winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day, and complex software engineering and enterprise work demand the same long horizon coding and coordination.

That is the real benchmark here. Most agent demos still last minutes. This run lasted more than a day, survived six server restarts, recovered from hundreds of deaths and kept four agents aligned on one goal through chat and shared notes. The habits that made it work, namely splitting roles, writing down lessons, building small reusable tools and correcting false reports, are exactly what companies want from coding agents in long projects.

On the safety side, the setup is a useful template too. The agents' game programs could control only their own character, could not start processes or write files, and never used a server command. Every program, message and event was recorded, an independent Observer agent checked the players' claims, and researchers could pause or stop the run at any time from a live dashboard. They had no reason to intervene.

It is also a reminder of how fast this field moves. A few years ago, a goal like this needed purpose built systems trained on enormous amounts of gameplay. Now a general model, given a library and one sentence, finishes the game for the price of a decent laptop.

Sources

• Primary source: Faldor, Singh, Raileanu, "A team of Gemini 4 Argon agents beats Minecraft", arXiv 2610.22319, 15 October 2026
• The Decoder, 18 October 2026