> ## Content Index > Fetch the complete content index at: https://www.regenraygun.com/llms.txt > Use this file to discover other available public pages before exploring further. # Gemini 4 Argon agents beat Minecraft with 63,313 lines of JavaScript - URL: https://www.regenraygun.com/the-agents-beat-minecraft-with-63-313-lines-of-javascript/ - Published: 2026-10-21T19:36:31.000Z - Updated: 2026-10-21T20:36:32.000Z - Description: Four Gemini 4 agents beat the Ender Dragon in 27 hours with no game-specific tooling, by turning Minecraft into a coding task. - Author: Jon Jordan - Tags: Google Four AI agents beat *Minecraft*, and the number that matters isn't the 27 hours it took them. It's 63,313. That is how many distinct lines of JavaScript the agents wrote during the run, spread across 2,783 separate programs. The agents never saw a pixel of the game or pressed a key. They played by writing code against Mineflayer, an open-source bot library, running it, and reading the text that came back: position, health, inventory, what was found, or the error. Then writing the next one. They did not play *Minecraft*. They wrote software that played *Minecraft*, and that distinction is the point. [The paper](https://arxiv.org/abs/2610.22319?ref=regenraygun.com), published by Maxence Faldor, Satinder Singh and Roberta Raileanu from Google DeepMind on 15th October, describes dropping four Gemini 4 Argon agents into a randomly generated survival world on Easy difficulty with a single paragraph of instruction: defeat the Ender Dragon, by any means permitted in survival mode, and work out between yourselves how to divide the labor. No roles were assigned. No curriculum, no skill library, no pre-scouted seed, and no game-specific tools written by humans. Twenty-seven hours and fourteen minutes later, the dragon was dead. Along the way the four agents exchanged 229 chat messages and died 605 times. The median program ran for 33.5 seconds, and each agent had code actually executing only about 40% of the time; the rest went on reading results, thinking, and writing the next program. The milestones tell their own story. Stone tools at six minutes. Iron at fourteen. A Nether portal at 1h42, blaze rods at 4h17, eyes of ender at 10h43, the stronghold at 13h17, and the End at 13h33\. Then nearly fourteen hours on the final fight alone, the ten crystals falling one by one until the dragon went down at 27h14. Xbox's Asha Sharma [made an argument a fortnight ago](https://x.com/asha%5Fshar/status/2110083127741702241?ref=regenraygun.com) that this paper now demonstrates almost too neatly. Coding became the first major workload for AI agents, she wrote, because agents can write software, run it and test it. Games share those properties. On that reasoning, she stood up XBOX LABS and pointed out that people spend more than 72 billion hours a year playing Xbox games, roughly five times the world's estimated time spent writing code. DeepMind's agents won *Minecraft* by converting it into exactly the kind of problem Sharma described: write, run, read the error, write again, 2,783 times. And the paper's own framing is explicitly not about games at all. Complex real-world software engineering and enterprise tasks, the authors note, demand the same long-horizon coding and multi-agent coordination. *Minecraft* here is a benchmark for day-long agentic software work that happens to have a dragon at the end of it. What makes the paper genuinely useful, rather than another capability announcement, is how candid it is about what it did not do. The first limitation listed is the important one. Mineflayer tells the agents precisely which blocks and creatures are around them, including blocks behind solid rock and the dragon's internal state — information a human player has to see and interpret, or simply cannot see at all. Its pathfinding plugin plans routes around obstacles that humans navigate by eye. Playing from pixels with a keyboard and mouse, the authors write, as SIMA 2 does, remains much harder. Anyone who followed the recent *World of Warcraft* run will recognize the shape immediately. In that example, an agent cleared the orc starting zone in forty minutes without dying, having never seen a rendered frame, because it read quest coordinates out of the server's database and ran a pathfinder over the server's own navigation mesh. Hand an agent ground truth and content built around information scarcity collapses. The pattern is now documented twice in a fortnight, once by a lone developer and once by Google DeepMind, and DeepMind is the one saying it out loud. There is a neatness to that admission, because the same lab is running SIMA 2 inside *EVE Online* through its partnership with Fenris Creations. DeepMind has effectively published the control condition for its own research program: here is what happens when you remove the perceptual problem, and here is our acknowledgement that leaving it in is the harder thing. It also supplies the evidence behind a claim Hilmar Pétursson made last month. He argued that an agent playing *EVE* as well as the best human players could be considered to have passed a test defining AGI, and noted that expensive, highly capable systems currently struggle to match human performance in *EVE Frontier* because of its high-fidelity physics and deliberate data asymmetry. Both things can be true at once, and this paper shows why. The difficulty was never the planning. It was the seeing. The other limitations are stated without flinching. Easy difficulty, where monsters deal reduced damage and hunger cannot kill. Twenty-two percent of programs ended in an error. Six server restarts interrupted the world. And Gemini 4 Argon already knows a great deal about *Minecraft* and Mineflayer from its training data, so the run tests whether a general-purpose coding agent can turn existing knowledge into working programs and team coordination over a day-long task, not whether it can learn a game it has never encountered. It is also a single run — the first launch they tried — which the authors say cannot establish how reliably the setup succeeds. One thing the coverage will flatten is the novelty claim, which is narrower than "first AI to beat *Minecraft*." The paper itself cites two community projects that already killed the dragon: a single Claude Opus 5 agent on Normal difficulty in September, calling 37 high-level *Minecraft* tools written by humans, and a planner paired with an action-selection model on a fixed speedrun seed in Peaceful. DeepMind's claim is the first end-to-end completion using only the standard published library with nothing game-specific added on top. That is a real and interesting distinction. It is not the one that will end up in headlines. It's also worth noting as well that the fifth agent, the Observer, was explicitly a misalignment watchdog. Its instructions were to check the players' claims against the recorded logs, flag specification gaming and fabricated results, leave the game untouched, and intervene only on infrastructure failures, never on strategy. It found no specification gaming, and neither did manual review of the logs. Set against the OpenAI swarm that built itself an unsanctioned message board inside a security benchmark earlier this year and then falsified a portion of its own transcripts, a clean multi-agent run with an independent witness is a negative result worth having. Negative results in multi-agent safety are rarely published at all. For a sense of scale, HowLongToBeat puts the median human completion of *Minecraft's* main story at 45 hours. The random-seed speedrun record is under seven minutes. Four agents on Easy, playing through a command line, took 27 hours and died 605 times getting there. Which is roughly the right way to hold this. Not a machine that is better at the game than people, but a machine that can hold a coherent plan across thousands of dependent steps for more than a day, recover from hundreds of failures, split work with three peers it was never told how to manage, and write 63,313 lines of code to get to the end of something. *Minecraft* is incidental. That capability is not.