4 Gemini 4 Argon Agents Beat Minecraft in 27 Hours, No Humans

By Muzamil Ahmad · Updated 2026-10-19 · Chatbots

#chatbots

Save

Beating Minecraft used to be a research project of its own. Teams spent years training specialist AI systems that could do one thing: play this one game. Last week, Google DeepMind published a paper showing that era is over. Four copies of Gemini 4 Argon, Google's new flagship model, were dropped into a randomly generated Minecraft survival world with a single goal: kill the Ender Dragon, the game's final boss. Nobody gave them roles. Nobody handed them a strategy. Nobody wrote them a single game-specific tool.

Twenty-seven hours and fourteen minutes later, the dragon was dead. The agents had written 2,783 JavaScript programs to control their characters, swapped 229 chat messages, died 605 times, and, according to the researchers, completed Minecraft end to end autonomously for the first time. DeepMind shared the result on X on October 16, and the paper (by Maxence Faldor, Satinder Singh and Roberta Raileanu) landed on arXiv on October 15. Here is what actually happened, how the agents pulled it off, what the run cost, and the caveats the headlines skip.

What Actually Happened

The setup was deliberately unfair, in the agents' disfavor. The world was randomly generated, so nothing was pre-scouted. The game ran in survival mode, where you start with nothing, gather every resource yourself and can die in a dozen ways. Cheats were off. The only software connecting the model to the game was Mineflayer, a standard open-source library anyone can use to build Minecraft bots. The agents saw the world as text: health, inventory, nearby blocks and entities, all described in words. Everything else, the planning, the prioritizing, the recovering from disaster, had to come from the model itself.

The paper's headline claim is specific and worth quoting precisely: "To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans." Earlier Minecraft AI projects either trained a model purely to play the game, or leaned on human-written helper code for the hard parts. This team did neither. Engadget's write-up puts it plainly: the same general-purpose coding agents Google sells for software work were pointed at a game, and they finished it.

The researchers could watch the game live and could pause or halt it at any moment. They never needed to. The agents divided the work themselves within the first stretch of the run, and the division of labor that emerged reads like a human guild splitting jobs before a raid:

• janesick became the toolmaker, crafting the pickaxes, weapons and gear the team needed.
• gyldenia handled food and supplies, keeping four perpetually dying agents fed.
• auratian lit a key portal on the path to the End dimension.
• bulicame built Enderman shelters for the final push and landed the killing blow on the dragon.

None of those roles were assigned. They emerged from chat negotiation between the agents, which is the single most interesting detail in the whole paper.

How They Did It: Code, Chat and Shared Memory

The agents did not "play" Minecraft the way a person does, twitching at a keyboard. They played it the way programmers do anything: by writing code, running it, reading the error and trying again.

The code loop. Every action, mining a block, crafting a sword, walking to coordinates, was a JavaScript program written on the spot through Mineflayer, executed, and judged by its result. When a program failed, and 22% of them did, the agent read the failure, revised the code and retried. That write-run-debug loop, repeated 2,783 times, is what carried the team from punching trees to killing a dragon. The 605 deaths are part of the same story: death in Minecraft survival is cheap information, and the agents treated it that way, respawning, walking back and collecting their dropped gear, sometimes with a teammate's help.

The chat channel. The 229 messages were not decoration. Agents warned each other about threats, passed on coordinates, shared lessons ("this approach failed, do it this way") and coordinated who was doing what. When one agent learned something the hard way, the others did not have to die learning it again.

The shared memory. Mistakes were logged in a memory all four agents could read. That matters more than it sounds. A single agent with no memory repeats its errors forever. Four agents with a shared error log compound their learning: every failure made the whole team smarter. Our guide to Google's Gemini Agent covers how this pattern, persistent memory plus tool use plus self-correction, is becoming the standard recipe for agents that run longer than a single chat reply.

The Observer. A fifth Argon agent monitored the session as an independent witness, checking for cheating and publishing live and summary reports. That is a genuinely new wrinkle: an AI auditing other AIs, in real time, on a task too long for a human to watch end to end.

Why This Matters Beyond Minecraft

It is tempting to file this under "fun demo" and move on. That would miss the point, for three reasons.

First, the generalist beat the specialists. A few years ago, beating a game like Minecraft required a bespoke system: custom training, custom perception, months of compute aimed at one environment. AlphaGo and AlphaStar, DeepMind's famous game champions, were masterpieces of exactly that approach. The Argon agents are the opposite: one general model, no game-specific training, no custom weights, succeeding through reasoning and code. The skill being demonstrated is not "playing Minecraft." It is "walk into an unfamiliar system with a long goal, figure out the rules, and finish the job."

Second, the horizon is the story. "Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day," the team wrote. Long-horizon reliability is the exact bottleneck holding back agent deployments in real work. An agent that can hold a plan together for 27 hours in a chaotic sandbox, recovering from 605 deaths and hundreds of failed programs without drifting off-task, is demonstrating the same capability a business wants from an agent handling a multi-day research, coding or operations task. Minecraft is just a conveniently measurable stand-in.

Third, teamwork emerged instead of being programmed. Nobody scripted "one of you make tools, one of you handle food." The roles fell out of the agents negotiating through chat and noticing what needed doing. Multi-agent coordination is where most real-world agent projects are heading, and this is a public, verifiable example of it working for more than a day straight.

None of this means your next software project should be staffed by game-playing bots. It means the underlying model, the one we benchmarked in our full Gemini 4 Argon guide, can sustain agentic work far longer than the short demos the industry usually shows.

What the Run Cost

DeepMind did something refreshingly honest: it published the token bill.

The session used 1.35 billion input tokens across the four players plus the Observer. Priced at Argon's introductory rates, the full session comes to roughly $810. At the standard rates, it is roughly $1,620, the figure most coverage led with. The dragon-killing run itself, excluding the Observer and the extra day of activity after the victory, was cheaper: around $520 at launch pricing and about $1,040 at standard pricing.

Why two prices? Because Argon is in a promotional window. As we detailed in the Argon pricing breakdown, introductory API rates of $2 per million input tokens and $10 per million output tokens run only through November 30, 2026. On December 1 the rates double to $4 and $20. So the same Minecraft run would cost noticeably more if repeated in December, and the "$1,620 dragon" headline is really the post-promotion price of the full session.

The honest read on cost: $810 to $1,620 for one dragon kill is not a business case. It is a benchmark receipt. What it proves is that a frontier model can convert roughly a thousand dollars of inference into a full day of coordinated, self-directed work. Whether that trade is worth it depends entirely on the value of the work, which in a game is zero and in, say, a week-long code migration is very much not.

Honest Caveats

The headlines are real, but they compress a lot. Here is what to keep in mind before citing this as proof that agents have "solved" long tasks.

• One run, no repeat rate. The paper describes a single successful run. It does not tell us how often the team fails if you restart the experiment, and 22% of individual programs failing hints at how bumpy the path was.
• Easy difficulty, and the model already "knows" Minecraft. The run used easy mode, and Gemini's training data certainly includes enormous amounts of Minecraft guides, wikis and videos. The agents were not discovering fire from first principles; they were applying memorized game knowledge through fresh reasoning. Still impressive, differently impressive.
• Humans built the arena. Researchers chose the library, set the goal, could watch live and could halt the run at any time. "No human help" means no human gameplay input, not no human involvement.
• Slow by human standards, fast by agent standards. A typical player finishes Minecraft's main story in a median of about 45 hours (per HowLongToBeat), and speedrunners on random worlds do it in under seven minutes. The agents took 27 hours of wall-clock time and roughly 81 days of in-game time, with six server restarts along the way. The achievement is autonomy and endurance, not speed.
• It is a sandbox, not the real world. Minecraft has perfect information available through the API, instant resets after death and no consequences for failure. Real work offers none of those mercies.

What This Means If You Use Gemini

For regular Gemini users, nothing changes today: this was a research demonstration run through the developer stack, not a feature rolling into the Gemini app. For developers and teams evaluating Argon, the takeaway is narrower and more useful. Argon's pitch, a million-token context, strong coding and an agentic toolkit, now has a public stress test behind it: the model sustained a coherent multi-agent plan for over a day and finished. If you are weighing it against GPT-6 or Claude for long-running agent work, our Gemini 4 Argon guide has the benchmarks, access routes and the December price step laid out, and the introductory pricing window closes November 30.

The Bottom Line

Four general-purpose AI agents, given nothing but a goal and a standard bot library, organized themselves into a team, wrote nearly 2,800 programs, died 605 times, learned from every death in a shared notebook, and killed the Ender Dragon. A decade ago that sentence would have described a moonshot research program. In October 2026 it describes a side experiment run on a model you can rent by the token, with the receipt ($810 to $1,620) published for anyone to check. The caveats are real: one run, easy mode, a game the model already knows well. But as a demonstration of where agentic AI actually is, long-horizon, self-correcting and quietly collaborative, this is the clearest public evidence yet.