Bare hands, a random seed, and a one-paragraph objective. DeepMind's Maxence Faldor, Satinder Singh, and Roberta Raileanu put four Gemini 4 Argon coding agents into a sandboxed Minecraft survival world and left them alone until the dragon fell.
The paper (arXiv:2610.22319) frames the run as a long-horizon coding and multi-agent coordination test: the same skills enterprise software agents need when work stretches across thousands of dependent steps. No roles were assigned. No game logic was written by humans. Researchers watched a live dashboard and could pause the session; they say they had no cause to intervene.
First end-to-end autonomous Minecraft clear on stock Mineflayer
Authors claim, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted seeds or other human-written game-specific tools. Each of four players was a general-purpose Gemini 4 Argon coding agent on its own cloud machine. Agents acted by writing short JavaScript programs that a Mineflayer bot executed; they read text returns for position, health, inventory, findings, and errors. Median program runtime was 33.5 seconds. Players shared in-game chat and a team memory. A fifth Observer agent witnessed the session, never played, and flagged no specification gaming.
Minecraft Java Edition 1.20.4 ran in survival on Easy in a randomly generated world, cheats disabled, no human input during the run. Beyond a brief note that they were a team of four, game instructions were a single paragraph: defeat the Ender Dragon by any means lawful in survival, and decide how to share the work. General coding instructions asked agents to verify actions, build reusable tools, record lessons, and report honestly.
Twenty-seven hours, 2,783 programs, 605 deaths
From first program to dragon defeat took 27 hours 14 minutes, spanning about 81 in-game days and six world-server restarts (about 20 minutes paused). The four agents wrote and ran 2,783 programs before the win, exchanged 229 chat messages, and died 605 times. The server log records bulicame's advancement Free the End at the finish; none of the programs used a server command.
Roles emerged in chat within the first quarter hour. Within two hours the team had iron tools, diamonds, and a Nether portal. Nether work produced blaze powder and ender pearls; losses to lava and Ghasts became posted rules. Stronghold triangulation from two Eye-of-Ender throws landed 43 metres from the portal room. The End fight killed the team 69 times; healing crystals on ten obsidian pillars kept the dragon topped up until agents climbed with dirt and punched them by hand. Thirty-nine minutes after the last crystal fell, the dragon died with bulicame at 2.5 of 20 health.
Token counters for the four players through the win estimate about 1.35 billion input tokens (91% from context cache) and 16 million output tokens across roughly 24,000 model calls. At Gemini 4 Argon introductory list price that is about $520; at standard list about $1,040. Including Observer and post-win activity, the whole session sits near $810 on introductory pricing.
Easy mode, symbolic senses, and a packed prior
Comparison fairness is narrow. Prior Minecraft agent lines either learn policies from gameplay (MineRL, VPT, DreamerV3, SIMA) or pair language models with Minecraft-specific planners, curricula, and skill libraries (Voyager and related systems). Community projects have reported dragon kills with high-level tool packs or fixed speedrun seeds. This paper's contrast is the interface: stock Mineflayer plus pathfinding, one-paragraph objective, general coding instructions, random Easy seed, team chat.
Easy difficulty lowers monster damage and prevents hunger death. Mineflayer exposes exact nearby blocks and creatures (including through solid rock) and the dragon's internal behaviour phase, information a pixel-and-keyboard player must infer or cannot see. Pathfinding plans routes that humans plot by eye. Gemini 4 Argon already carries substantial Minecraft and Mineflayer knowledge from training. The measured skill is turning that knowledge into working programs and day-long coordination, measured once.
Reliability across seeds and Hard difficulty remain unsettled
The paper describes a single run (the first launch attempted), so success rate across seeds is unsettled. Harder difficulties, pixel-based control, withheld Minecraft priors, and multi-run statistical claims are outside the reported design. Twenty-two percent of programs ended in error; infrastructure restarts interrupted play six times. Cost figures are list-price estimates from session counters rather than a serving invoice.
Paper, logs claim, and open texture credits
Primary artefact: arXiv:2610.22319, "A team of Gemini 4 Argon agents beats Minecraft." Authors state that every program, output, chat message, and the full server log were recorded, with the win timestamped in the server log. In-game figures use the open-source Pixel Perfection texture pack. Gemini 4 Argon weights and the exact cloud agent harness are product surfaces; the paper does not ship a public reproduction bundle as a companion release in the abstract.
Borrow the coordination pattern, keep the caveats
Teams building multi-agent coding systems can treat the run as evidence that shared memory, honest correction, item recovery, and self-authored helper libraries can sustain a day-scale goal when the environment API is rich. Product managers should refuse to translate "beat Minecraft" into "ready for unsupervised production ops" without Hard-mode, multi-seed, and pixel-control follow-ups. Eval designers can copy the Observer pattern: an independent witness tasked to catch specification gaming while barred from helping.
MineDojo's moonshot, finished on Easy with code as policy
MineDojo listed defeating the Ender Dragon from bare hands in a fresh world among its moonshot goals. This paper sits in the code-as-policy and multi-agent Minecraft line rather than the pixel-control SIMA line, and it differs from recent community dragon kills mainly by refusing game-specific human tools and pre-scouted seeds. The next papers in the line will need repeated runs and harder perception stacks before the result travels as a reliability claim.