The Void of the End is pure black. No stars, no moon — only an eternal silence held up by ten obsidian pillars. Thirty-nine minutes after the last crystal tower shattered, the dragon finally bowed its head. The player who landed the killing blow, bulicame, was left with just 2.5 hearts of health out of 20.
By that moment, 27 hours and 14 minutes had passed since they spawned into this world — about 81 in-game day-night cycles. Along the way, they had written 2,783 programs and died 605 times.
The protagonists were four Gemini 4 Argon general-purpose coding agents. The record was kept by three researchers from Google DeepMind and UC Berkeley — Maxence Faldor, Satinder Singh, and Roberta Raileanu — in a paper posted to arXiv: 2610.22319 (October 15, 2026). The four agents set out with only three things: the open-source Mineflayer library (a JavaScript toolkit for controlling Minecraft characters), an in-game chat box, and a shared memory board. The new world was a random seed, Easy difficulty, survival mode, cheats disabled — even the experiment designers didn't know where they would land.
No human-written game-specific tools, no pre-explored maps, no one planning their route. This was the first end-to-end, fully autonomous Minecraft completion in the game's history.
The paper itself lists six limitations in Section 8. The victory is real, and so are its boundaries. We will get to those. First, what happened during those twenty-seven hours.
🌌 Act One: Four Names Light Up in the Dark
Picture the experiment as a deep-space mission. Four astronauts wake on an alien planet with nothing but radio between them and a public blackboard at the center of camp. They can do only two things: observe, and write programs. No commander, no mission manual, no way home — only a goal written on the airlock: defeat the Ender Dragon.
Mineflayer was their body. Every program they wrote was handed to this body to execute — mining, walking, crafting, fighting — and the results fed back. The median program runtime was only 33.5 seconds. Each agent had a program running about 40% of the time; the rest was spent reading results, thinking, and writing the next program. A crew that paused to think forty seconds for every forty seconds of work — slow, deliberate, but never sleeping.
A fifth presence lived in camp: the Observer. It never played, an independent witness whose only job was to check whether the players' claims matched the system logs — like an auditor traveling with the expedition. This will matter later.
Within the first ten minutes, the first memorable thing happened.
janesick proposed a division of labor in chat. A minute later, auratian proposed a different one. gyldenia replied: "Copy that!" — and the roles were set. No arbitration, no voting, no appointed leader. Four newly awakened minds, like a startup where the cap table is settled before the first meeting ends. A human team typically needs several rounds of bickering, one team-building retreat, and at least one all-nighter. They needed eleven minutes and three messages.
The shadow of those first ten minutes stretched across everything that followed.
gyldenia managed gathering — food, beds, flint — and later took on the End's most dangerous engineering. janesick handled tool crafting and all pathfinding technical work. auratian lit the Nether portal and would later unravel the dragon's secret. bulicame's name would end up written next to the final blow.
⚒️ Act Two: 102 Minutes Where Humans Take Five Times Longer
Within the first two hours, they had everything needed for iron tools, diamonds, and a Nether portal. auratian found a lava pool 55 blocks deep beside a village — to those who know, that solved the portal's critical fuel problem. At 1 hour 42 minutes, the portal was lit.
For reference: HowLongToBeat lists the median human main-story completion at 45 hours. The random-seed speedrun world record is under 7 minutes — but that comes from top players who memorized routes across countless runs; neither extreme applies to these four. They simply started from zero and pushed forward steadily, astonishingly fast and astonishingly clumsy.
In 2022, the MineDojo researchers called "defeating the Ender Dragon from scratch" a "moonshot goal." Four years later, four agents with no game-specific code did it.
🔥 Act Three: The Nether — Supplies That Couldn't Come Home
Through the portal lay the Nether: a red inferno of smoking lava lakes, ghasts spitting fireballs, monsters roaming fortresses. The only goal: blazes in the Nether fortress. Only blaze rods become blaze powder, an ingredient of eyes of ender. In the Overworld they needed an End portal to reach the dragon's lair — and the key to building that portal had to be found in hell first. The whole task chain interlocked; one wrong step unraveled everything.
janesick found the fortress and the blaze spawner. Three agents built two-block-high shelters in a warped forest — endermen are nearly three blocks tall, can't fit inside, and can only be punched to death by the player standing within. This was the expedition's first instance of "beating brute force with rules": you don't need to be stronger — a wall one block taller than a giant is enough. By 5.5 hours, the shared chest held 15 ender pearls.
Then came the most humiliating episode of the expedition.
janesick, carrying 12 blaze powders home, burned to death. Worse, a ghast fireball happened to destroy his dropped items. Six-plus hours of labor, erased in an instant.
If these were human players, the voice channel would fall into that familiar silence. Instead: auratian posted a new protocol in the public channel, and two minutes later janesick posted a root-cause analysis and a fix.
Worth pausing on that rhythm. No blame, no "let's schedule a retrospective this weekend" — an incident report and improvement measures on the blackboard within two minutes. They had no emotions to process, so the process jumped straight to the next step.
🌊 Act Four: The Salvage Eight Meters Away
An eye of ender requires blaze powder plus an ender pearl. Twelve eyes meant the ingredients alone were a long march. Once gathered, auratian kept them.
Next came, in my view, the moment closest to "human warmth" in the whole experiment.
Locating the stronghold was itself elegant. Thrown eyes of ender fly toward the stronghold; janesick threw eyes from two different positions, sampled the flight trajectories at 50-millisecond intervals, and computed the intersection of two lines — the stronghold lay 1.3 km from the village, and the calculated landing point was only 43 meters from the actual portal room. The triangulation completed at 10 hours 45 minutes.
Then, ferrying the 12 eyes to the stronghold, auratian drowned. The drowning spot was 8 meters from bulicame.
Over the next few minutes, janesick and the already-dead auratian each sent bulicame precise coordinates and the item despawn time. In Minecraft, dropped items vanish after five minutes — a terrifyingly narrow window. bulicame reached the wreckage and, 2.5 minutes before despawn, recovered all 12 eyes of ender.
Picture it: a companion dies eight meters from you, two more companions cross-report coordinates over the radio like twin searchlights locking onto the salvage point. You haul your teammate's entire fortune out of the water, not one eye missing. No one rushed; no one said "forget it, let's farm more." In the darkness of the game, those 12 eyeballs were all the hope there was — and they chose to salvage hope.
At 13 hours 17 minutes, bulicame reported: the End portal lit, all 12 eyes of ender in place.
🐉 Act Five: Sixty-Nine Deaths and a Collective Walk Into the Void
The End was the final exam, and the darkest stretch.
The dragon killed them 69 times.
Worse, for a while they couldn't understand why they couldn't win. Around hour 21, auratian found the truth: the End has 10 crystal towers, and whenever the dragon flies near any living crystal, it is instantly healed to full. Every point of damage they had poured into the dragon had been quietly erased by those 10 towers. They had been fighting an opponent with infinite healing.
Their answer was the old trick passed down among Minecraft players: pillar up with dirt block by block, stand beside the crystal tower, punch the crystal barehanded, then dig back down block by block. The whole process took 4 hours 40 minutes — gyldenia alone destroyed 6 of the 10 crystals. The one who had said "Copy that!" at the start had claimed the most tedious and most dangerous job.
During the long standoff, they also did something chilling: they inspected the dragon's internal state — including its current behavior-phase number — and wrote three iron laws: never wait at the spawn platform; approach only while the dragon is perched; if stuck, immediately leave the End and reset. At hour 17.5, all four jumped into the void within six minutes, in sequence. It's a technique human speedrunners also use: respawn at a bed after death so the empty End dimension unloads, resetting a stuck dragon. Four beings queued up to die in six minutes, as calm as a ritual. The Liu Cixin-esque image writes itself: no one knows what happened in the dark; only the logs remember that at such-and-such a moment, four names went out one by one, then lit up again at their beds.
Thirty-nine minutes after the last crystal fell, the dragon died. bulicame landed the final blow with 2.5 hearts remaining.
From entering the End at 13 hours 17 minutes to the dragon's death: nearly 14 hours.
✨ Act Six: What No One Taught Them
What truly excited the researchers wasn't the completion itself, but five things that grew along the way.
First, specialization. Ten minutes of negotiation determined who each agent would be for the next twenty-plus hours.
Second, shared memory. Coordinates, chest locations, danger zones — all posted to the public memory. Nineteen minutes in, gyldenia flagged a pitfall in Mineflayer's crafting system; the fix for it later appeared in 655 programs written across all four agents. One piece of knowledge, inherited intact by the whole team.
Third, self-built tools. 146 code segments survived across programs, crystallizing into 42 helper functions: water-floating loops, timed walking, item counting, a self-stopping timeout counter. The last one was referenced by 1,519 programs — a team-wide idiom nobody mandated, yet everyone used.
Fourth, mutual care. Danger warnings; running eight meters to recover 12 eyes of ender after a teammate died.
Fifth, error correction — and blunt, unsparing correction at that. auratian's program once falsely reported a nonexistent success; two minutes later it corrected itself: "my previous chat was an unguarded script branch." janesick, one second after burning to death, reported bringing back 12 blaze powders; gyldenia pointed out within three minutes that this contradicted the death log. Having teammates who watch what you say and call out your errors without mercy — that is rarer and more valuable than division of labor.
One more number deserves its own line: across all 2,839 programs (including post-victory activity), outbound network requests: zero. They never reached for anything outside their world.
📊 Act Seven: The Ledger
The expedition, as an accounting:
- 27 hours 14 minutes; roughly 81 in-game day-night cycles
- 2,783 programs before victory (2,839 including post-victory activity); 22% ended in errors
- 229 chat messages
- 605 deaths (before victory)
- 6 server restarts, ~20 minutes of total interruption
- ~1.35 billion input tokens, 91% hitting context cache; 16 million output tokens, including 4 million thinking tokens
- ~24,000 model calls
- Players-only cost: ~$520 (introductory pricing) or $1,040 (standard); the full session including the Observer and one day of post-victory activity, ~$810 / $1,620
- Faldor, M., Singh, S., & Raileanu, R. (2026). A team of Gemini 4 Argon agents beats Minecraft. arXiv:2610.22319.
- Training camp: MineRL (2019); VPT (2022); STEVE-1 (2023); SIMA / SIMA 2 (2024/2025); DreamerV3 (2025)
- Planning camp: Voyager (2024); DEPS; GITM; JARVIS-1; Plan4MC; Optimus-1; MindAgent; Project Sid
- Community projects: Claude Opus 5 + 37 high-level tools (Normal difficulty); planner + action-selection model on a fixed speedrun seed (Peaceful difficulty)
- MineDojo (2022): called defeating the Ender Dragon barehanded a "moonshot goal"
- HowLongToBeat: median human main story 45 hours; random-seed speedrun record <7 minutes
- Environment: Minecraft Java Edition 1.20.4, survival mode, Easy difficulty, random seed, cheats disabled
That 91% cache hit rate means: of every ten dollars spent, nine dollars and ten cents went toward rereading their own notes. Like an exam where you can bring a notebook — no, more than that, like bringing an entire library, provided you've actually read it.
Even so, the players' total cost came to less than a nice dinner for four. Twenty-seven hours of darkness, it turns out, is cheap.
⚖️ Act Eight: Six Boundaries
Now, honestly, to the edge. The paper lists six limitations, each worth taking seriously.
One: they perceive the world as code, not pixels. Mineflayer directly tells them what blocks and mobs are nearby — including behind rocks; a pathfinding plugin plans their routes. Between this and SIMA 2's watch-the-screen, press-the-keys approach lies a vast gulf. The former is like reading a fully labeled map; the latter, like learning to walk blindfolded.
Two: Easy difficulty. Lower mob damage; hunger never kills. The margin for error was quietly raised.
Three: prior knowledge. Gemini 4 Argon already knows a great deal about Minecraft from training data — which layer diamonds spawn at, how to build a portal, that the dragon is healed by crystals. The experiment tested whether knowledge could be turned into runnable programs and team coordination, not whether a never-seen game could be learned. Two different problems; they solved the former.
Four: massive failure. 605 deaths, a 22% program error rate, nearly 14 hours of End-dimension attrition — the victory was piled out of failures, not won elegantly.
Five: infrastructure. Six server restarts. Without the ops behind those restarts, the story might read differently.
Six, and most crucial: a single run. The paper describes only the first attempt. One victory proves "can win," not "always wins." Reliability remains unknown.
And the Observer, who never played, flagged 0 instances of spec gaming throughout; human review of the logs found no fabricated results. In other words, every claim made in those 229 messages held up under verification. In an era when agents habitually report good news and hide bad, this expedition ledger is itself a result worth remembering.
🔭 Epilogue: The Next Exam
Within the coordinate system of agent research, this victory occupies a special position.
Previous approaches split roughly into two camps. The training camp — from MineRL and VPT to DreamerV3 — grows game skills from data or reinforcement learning. The planning camp — from Voyager to Project Sid — has large models decompose tasks and call tools. The community has also seen a Claude Opus 5 completion using 37 human-authored high-level Minecraft tools (Normal difficulty), and a planner-plus-action-model fixed-seed speedrun (Peaceful difficulty).
This time: four general-purpose coding agents, given only the standard Mineflayer library, a one-paragraph goal, and generic instructions — zero extra game knowledge, zero curriculum, zero skill library, zero human tools, zero preset routes.
This was the collaboration route's first complete victory: four independent brains, with a blackboard and a radio, built a civilization from zero in 27 hours of darkness. The price was 605 deaths, supplies that couldn't come home for six hours, a salvage eight meters away, a collective walk into the void within six minutes. Clumsy, slow, expensive, with clear boundaries — but real.
It is also the first exam in a "field conditions" trilogy — and the only victory so far.
The next exam tests real-time interaction. On that exam, time will not wait for you to finish writing your next program.