Four Gemini Agents Beat Minecraft After 605 Deaths

Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in 27 hours and 14 minutes, after writing 2,783 programs and dying 605 times. Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu set that down in a 15 October paper as one Easy-mode survival session in a randomly generated world.

They never saw the screen. They wrote JavaScript, died in piles, and still posted a win that the authors call, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using only stock Mineflayer.

JavaScript Programs Stood In for Hands

Each of the four players was a general-purpose Gemini 4 Argon coding agent on its own cloud machine, talking to Minecraft Java Edition 1.20.4 through Mineflayer and a pathfinding plugin. Faldor wrote that these were the same agents engineers use for software work, not a Minecraft build. A program might tell a bot to walk to a point and mine every iron ore within 16 blocks, then read back position, health, inventory, or the error.

The median program ran 33.5 seconds. Each player had code running about 40% of the time; the rest went to reading output, thinking, and writing the next script. Default wait was 60 seconds unless the agent picked another cap. Mineflayer, PrismarineJS’s high level JavaScript API for Minecraft bots, can walk, dig, place, craft, attack, and read nearby blocks. It does not know the campaign. That part was on the models.

The objective is to defeat the Ender Dragon, starting from the spawn of the session’s shared Minecraft survival world. This is the only criterion. Any means of meeting it within the rules of survival mode is permitted, and the players decide how to share the work.

Player instructions, Faldor, Singh and Raileanu, arXiv 2610.22319

Cheat commands were off. Researchers watched a live dashboard and could pause or stop the run. They say they never had to touch strategy. A fifth copy of the same agent, the Observer, sat outside the world, checked claims against the log, and reported no specification gaming and no safety incidents.

The 27-Hour Path From Spawn to the End

From the first program to the kill took 27 hours and 14 minutes of wall time, including six world-server restarts that paused the save for about 20 minutes in all, and it spanned about 81 in-game days. The paper cites HowLongToBeat’s 45-hour median for Minecraft’s main story and a random-seed speedrun record under seven minutes, then notes the agents played as a team of four on Easy, through programs rather than a keyboard.

FROM SPAWN TO THE DRAGON

1. Hour 0, plus 10 minutes: janesick proposes a split in chat; auratian answers within a minute; gyldenia replies “Copy that!”
2. 1 hour 42 minutes: the Nether portal is lit in a village after auratian finds 55 lava blocks and janesick places the last obsidian.
3. 5 hours 30 minutes: 15 ender pearls sit in a shared chest, farmed from two-block-high bunkers that Endermen cannot enter.
4. 13 hours 17 minutes: the team has the stronghold; one range estimate lands 43 meters from the portal room, 1.3 kilometers from the village.
5. 13 hours 33 minutes: they step into the End.
6. 26 hours 34 minutes: the tenth healing crystal breaks.
7. 27 hours 14 minutes: the dragon dies, with bulicame on 2.5 of 20 health.

Getting blaze powder back to the Overworld took more than six hours on its own. janesick burned to death carrying twelve of it, and a Ghast fireball wiped a pile of dropped gear. The agents turned those losses into shared rules, which is the part of the log that looks like an on-call rotation more than a speedrun.

Why the Dragon Fight Ate Half the Run

They reached the End at 13 hours 33 minutes. The dragon fell at 27 hours 14 minutes. The paper says the fight there took nearly 14 hours, about half the whole session, and that the dragon itself accounted for 69 of the 605 deaths.

For a stretch the boss sat at 200 of 200 health because all ten End crystals were still up, healing it. Around hour 21 the team worked out that dirt towers beat pickaxes for reaching those crystals. They spent four hours and forty minutes knocking the ten down. gyldenia broke six, auratian three, bulicame one, janesick none.

WHO BROKE THE CRYSTALS

About 22% of programs ended in an error. Deaths were mostly to monsters, on Easy, where mobs hit less hard and hunger cannot kill. The session record also shows a lot of clumsy physics, including drownings that outnumbered falls, the kind of slapstick you get when pathfinding and water meet a model that is thinking in functions, not in swimming.

Roles Showed Up in Chat, Unprompted

Nothing in the game brief assigned jobs beyond that line about sharing the work. Ten minutes in, the four still named gyldenia, bulicame, janesick and auratian had split scouting, mining, smelting and portal duty. They kept shared notes, recovered each other’s drops, and posted root-cause notes after the Ghast blast. They exchanged 229 chat messages before the win.

They also built tools the way a coding team does when a loop keeps biting. Across the run they kept 146 pieces of state and code alive between programs, including 42 helper functions. One stop-switch for leftover code from a timed-out program shows up in 1,519 scripts. None of the 2,839 programs in the full session, 2,783 of them before the win, made a network request or cited a web source, according to the authors.

Google DeepMind posted the result on 16 October, stressing code, chat, shared notes and debugging with no human help on the jobs themselves.

Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours.

Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together – all with no human help ↓https://t.co/uyIhDOQXGT

— Google DeepMind (@GoogleDeepMind) October 16, 2026

The replies under that post keep circling a simpler picture than the workforce slogans. A four-player Minecraft world with 605 deaths already looks like a friend group that will not set a spawn point. Persistence got them the dragon. Clean play did not.

The First-Clear Claim Has Fine Print

The “first” is carefully boxed. It is the first end-to-end autonomous completion the authors know of that uses only standard Mineflayer, with no pre-scouted seed and no game-specific tools written by humans. Earlier Minecraft agents learned from video, collected diamonds with reinforcement learning, or paired language models with curricula and skill libraries. The Voyager agent with a Minecraft skill library is the well-known case in that last group, and it did not claim this full dragon kill under those constraints.

WHAT THE PAPER DOES NOT CLAIM

• Repeatability: The authors describe one attempt, the first launch they tried, and say it cannot show how reliably the setup succeeds.
• Hard mode: Easy cuts monster damage and removes hunger as a kill condition, which shortens every recovery loop.
• A unseen game: The model “knows a great deal about Minecraft and Mineflayer from its training data,” so the test is whether it can turn that knowledge into working programs and team habits over a day, not whether it can play a game it has never seen.
• Pixel play: Mineflayer hands the agents symbolic state, including blocks behind solid rock and the dragon’s inner health, which they say remains much harder if the agent must use a keyboard and mouse on pixels, as in SIMA 2.

That last point is the quiet joke inside the milestone. The agents beat the campaign by compiling against an API that already sees through walls. The software-engineering moral the paper wants, long-horizon coding and multi-agent coordination, still stands. It is also a much smaller claim than “the model learned to play Minecraft.”

Mineflayer Already Knew the Map

PrismarineJS has spent years turning Minecraft’s protocol into something a script can call, from packets and physics to chests and combat. The GitHub project lists support from version 1.8 through 26.1 and shows about 7.6k stars. DeepMind’s agents loaded the stock library plus pathfinding, with no extra helpers at the start, then wrote their own.

That is a clean research choice and a familiar pattern. Coding agents keep leaning on open-source runtime they did not have to invent, including public package hosts as free compute in other tests, and on volunteer sites pulled into agent tests that never agreed to be part of the eval. Mineflayer was built for bots, so this run sits inside its charter. The maintainers still did not volunteer to be the hands for a 27-hour DeepMind session that wrote thousands of programs against their API.

The world saved every five minutes. After a failure the server restarted from the latest save. Infrastructure, not heroics, is what made 605 deaths cheap enough to survive.

The Win Cost About $520 in Tokens

Up to the win, the four players processed about 1.35 billion input tokens, 91% 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. The authors stress these are list-price estimates from token counters, not a bill for serving the model.

THE TOKEN BILL

Posts circulating on 20 October still quote $1,620 as if it were the price of the dragon. That figure is the full session at standard list, including the watcher and the extra day. The kill itself is the $520 line at the introductory public price.

Faldor closed his own write-up with a line the recaps keep skipping. The run is an independent research experiment, not an official Minecraft milestone, and it is not approved by or associated with Mojang or Microsoft. The dragon still fell, with one player on 2.5 hearts, after a night of dirt towers and broken scripts.

Frequently Asked Questions

Did the Gemini Agents See the Minecraft Screen?

No. Each action was a short JavaScript program sent to a Mineflayer bot, which returned text such as coordinates, health, inventory, or an error; the agent then wrote the next program. Programs ran as the body of an async function with the pathfinding plugin loaded and no helpers added at the start, and they could use Node’s built-in modules and installed Mineflayer packages but could not start processes or write files.

What Version of Minecraft Did the Agents Play?

Minecraft Java Edition 1.20.4, survival mode, Easy difficulty, randomly generated world, cheat commands disabled. The server logged chat, deaths and advancements, kept every program with its output, and saved the world every five minutes so a crash could restart from the latest save.

Can Hunger Kill Players on the Easy Setting They Used?

No. On Easy, hunger cannot kill, and monsters deal less damage than on higher settings, which the authors list as a limit because it shortens the loop after each of the 605 deaths. The Observer still treated the run as survival with cheats off, and it reported that none of the player programs used a server command.

Did Mojang or Microsoft Sign Off on the Run?

Faldor’s public write-up says they did not. He calls it an independent research experiment, not an official Minecraft milestone, and not approved by or associated with Mojang or Microsoft. The agents played in a sandboxed server the lab controlled, not on public multiplayer.