Four Gemini 4 Argon agents defeated the Ender Dragon in a randomly generated Minecraft world in 27 hours and 14 minutes, according to a Google DeepMind paper posted on arXiv on 15 October 2026. The paper puts the cost of the run up to the win at about $520 at Gemini 4 Argon’s introductory API list price, and $1,040 at the standard price.
The researchers write that no one assigned roles or wrote game logic: the agents wrote 2,783 programs, died 605 times and coordinated through the game’s chat and shared notes. The paper says it describes a single run on Easy difficulty, “the first launch we tried, which cannot show how reliably the setup succeeds.”
Business Pill · AI AGENTS
A Business Engineer explainer on AI agents and the move from answering questions to doing tasks. It explains the concept in general and says nothing about DeepMind’s paper.
The key insight: As we read it, the number that matters for businesses is not the game but the duration and the bill: the paper reports four agents holding one goal for 27 hours and 14 minutes across 2,783 programs, at a list-price estimate of about $520, while stressing that it describes a single run.
The paper, by Maxence Faldor, Satinder Singh and Roberta Raileanu of Google DeepMind, describes four players each driven by the company’s general-purpose Gemini 4 Argon coding agent. Each agent controlled a player through Mineflayer, an open-source JavaScript library for Minecraft bots, and acted by writing and running short programs and reading the text that came back.
According to the paper, the agents saw no pixels and pressed no keys. The game ran in survival mode on Easy in Minecraft Java Edition 1.20.4, with cheat commands disabled and no human input during the run.
Their only instruction about the game was one paragraph, which the paper quotes: “The objective is to defeat the Ender Dragon, starting from the spawn of the session’s shared Minecraft survival world.”
The authors write that winning “required the agents to keep a coherent plan across thousands of dependent steps for more than a day.” They add that complex real-world software engineering and enterprise tasks demand the same long-horizon coding and multi-agent coordination.
Estimated cost of the Gemini 4 Argon Minecraft run at API list prices, from section 6 of Google DeepMind’s paper (arXiv 2610.22319): about $520 up to the win at the introductory price ($1,040 at the standard price), and about $810 for the whole session including the Observer agent and the day after the win ($1,620 standard). The paper says these are estimates from list prices and token counters, not the cost of serving the model.
What Was Posted on X
The paper on how a team of Gemini 4 agents beat Minecraft is worth a look.
It is accessible example (partially because it is AI written) of spontaneous coordination among agents to accomplish a complicated goal & how their approach differs from humans/ https://arxiv.org/pdf/2610.22319
Ethan Mollick shared the DeepMind paper on X on 16 October, calling it an accessible example of spontaneous coordination among agents and noting that it is partly AI written. The piece quotes his post and is built on the paper itself.
What the Run Cost
The paper’s cost section says that 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.
At Gemini 4 Argon’s introductory list price that comes to about $520, or $1,040 at the standard price, the paper says. The whole session, including the Observer agent and the day of activity after the win, came to about $810, or $1,620 at the standard price.
The authors add: “These figures are estimates from list prices and the session’s token counters, not the cost of serving the model.”
By our arithmetic, $520 over 27 hours and 14 minutes is about $19 an hour, and about 19 cents for each of the 2,783 programs. This week we reported that Gemini 4 Argon’s introductory price doubles after 30 November.
How the Agents Split the Work
The paper says nothing in the instructions mentioned roles, beyond a line that the players decide how to share the work. It reports that each agent settled into one: gyldenia gathered supplies and later destroyed six of the ten crystals, janesick made tools, found the fortress and triangulated the stronghold, auratian lit the Nether portal, and bulicame landed the final blow in the End.
According to the paper, the team exchanged 229 messages before the win, 173 of which contained exact coordinates. It says the agents kept 42 helper functions alive between programs, and that a fix for a crafting quirk posted by one agent 19 minutes in appears in 655 programs written by all four.
The paper also describes corrections. It says auratian retracted a success its own program had announced two minutes after posting it, writing that “my previous chat was an unguarded script branch”.
A fifth agent, the Observer, followed the session to watch for cheating or specification gaming, the paper says, and neither it nor the researchers’ manual review of a sample of the logs found any.
What the Paper Says It Does Not Show
The authors list the limits themselves. They write that Mineflayer tells the agents exactly which blocks and creatures are around them, including blocks behind solid rock and the dragon’s internal state, and that “Playing from pixels with a keyboard and mouse, as SIMA 2 does, remains much harder.”
They note that on Easy, monsters deal less damage than on Normal or Hard and hunger cannot kill a player. They write that Gemini 4 Argon knows a great deal about Minecraft and Mineflayer from its training data, so the run tests turning that knowledge into working programs, “not whether it can play a game it has never seen.”
The paper counts failures along the way: “22% of programs ended in an error; the fight in the End took nearly 14 hours.” The world server restarted six times during the run.
The paper also cites two community projects that reported end-to-end dragon kills this year: a single Claude Opus 5 agent calling 37 high-level Minecraft tools on Normal difficulty, and a planner on a fixed speedrun seed in Peaceful. It says its setting differs mainly in what the agents were given.
What Mollick Said
Ethan Mollick shared the paper on X on 16 October. He wrote that it is an “accessible example (partially because it is AI written) of spontaneous coordination among agents to accomplish a complicated goal”.
The Structural Read
As we read it, the paper is a cost-per-task data point for long-horizon agent work: its authors tie the run to software engineering and enterprise tasks that demand long-horizon coding and multi-agent coordination, and estimate its cost from token counts at list prices rather than serving cost.
As we read it, the coordination layer did much of the work: the paper reports 229 messages, a shared memory, 42 helper functions kept alive between programs and a single crafting fix reused in 655 programs written by all four agents.
As we read it, the limits are part of the result: the paper says the agents read the game state as text through Mineflayer, played on Easy, drew on prior knowledge of the game and saw 22% of programs end in an error.
Google DeepMind, arXiv 2610.22319
“Complex real-world software engineering and enterprise tasks demand the same long-horizon coding and multi-agent coordination.”
Three Implications
TEAMS BUILDING AGENTS The paper reports that the agents’ general instructions asked them to verify each action, build reusable tools, record lessons in a shared memory and report outcomes honestly.
FINANCE AND PROCUREMENT The paper prices the run at about $520 at the introductory list price and $1,040 at the standard price, and says these are estimates, not the cost of serving the model.
RESEARCHERS AND EVALUATORS The paper says a fifth Observer agent watched for cheating or specification gaming, and that every program, message and game event was recorded.
The framework lists, among its claims, that “Orchestration, memory, tool use, and workflow execution have become the new sources of competitive advantage.”
As we read it, DeepMind’s run is a worked example of that list: the same general-purpose coding agent, given shared memory, chat and its own reusable tools, carried one goal through thousands of dependent steps. The framework was published in July 2026 and does not discuss this paper.
What Is Not Established
The paper describes one run. It does not report a success rate across repeated attempts, and the cost figures are list-price estimates rather than what it costGoogle to serve the model.
We read the arXiv paper in full on 17 October 2026, along with Mollick’s post. We did not contact Google DeepMind.
Business Engineer Framework
The Agentic Harness War
A Business Engineer analysis of why the agentic surface, the harness, is where value pools in the AI stack.
Google DeepMind says four Gemini 4 Argon agents completed Minecraft from a random world in 27 hours and 14 minutes, writing 2,783 programs, for about $520 at the introductory list price.
The paper itself stresses the limits: one run, Easy difficulty, text access to the game state through Mineflayer, and a model that already knows the game.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every figure on the run comes from Google DeepMind’s paper on arXiv (2610.22319, submitted 15 October 2026), which we read in full on 17 October 2026; the cost figures are the paper’s own list-price estimates. The per-hour and per-program figures are our arithmetic. Ethan Mollick’s comment is quoted from his post on X. We did not contact Google DeepMind. The reference to our earlier reporting links to our own piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
