Gemini 4 Argon agents beat Minecraft — the run cost $520
A Google DeepMind paper puts the first real number on autonomous agents — and a CNBC report finds the legal fine print hiding in your pocket.
Four Gemini 4 Argon agents took Minecraft from a randomly generated world to the Ender Dragon's death with no human touching the controls, in what the researchers call the first end-to-end autonomous completion of the game by AI agents. The team — Maxence Faldor, Satinder Singh and Roberta Raileanu — dropped the agents named janesick, gyldenia, auratian and bulicame into a standard survival server on Easy difficulty with one goal and nothing else: no assigned roles, no strategy, no human-written game tools. The agents divided the work themselves, warned each other of threats in in-game chat, recovered items from dead teammates and logged their mistakes in notes the others could read.
The run's numbers, from the paper posted to arXiv: 27 hours and 14 minutes from first program to dragon's defeat, 2,783 JavaScript programs written and debugged by the agents, 63,313 distinct lines of code, 229 teammate messages, and 605 deaths along the way. Each agent steered its player through Mineflayer, an open-source library that lets code drive a Minecraft character — no pixels, no keyboard. Each player had a program running only about 40% of the time; the rest was reading results, thinking and writing the next one. The whole session, including a fifth "Observer" agent that watched the record for cheating or rule-bending and found none, and a day of activity after the win, cost about $810 at Gemini 4 Argon's introductory API price, or roughly $1,620 at the standard price that takes effect in 2027. The win itself came to about $520 at introductory rates.
That cost-per-completion figure is the part enterprise buyers will care about. The agents' own argument is subtler: keeping one coherent plan alive across thousands of dependent steps for more than a day is exactly the skill software engineering and agentic-work benchmarks keep failing to measure. Minecraft has been beaten by AI before — but by systems trained from scratch on the game. This team's claim is that a general coding agent plus one standard library now suffices, with the same model that ships to everyone.
It's also not the frontier's first video-game headline this month: we watched Gemini 4 Argon play StarCraft last week, where the model got further than expected but never won.
The honest caveats are stacked in the paper's own limitations section. Mineflayer hands the agents information no human player gets — which blocks sit behind solid rock, the dragon's internal state — and its pathfinding plugin plans routes around obstacles a person would have to spot by eye; playing from pixels, as Google's SIMA 2 does, remains much harder. Easy difficulty means monsters hit softer and hunger can't kill. The agents already knew a great deal about Minecraft from training data, so the run tests whether a coding agent can convert that knowledge into working programs and team coordination over a day, not whether it can play a game it has never seen. And 22% of programs ended in an error, the fight in the End took nearly 14 hours, the world server restarted six times mid-run, and this was one attempt — a single run can't show how often the setup succeeds.
Even so, the researchers' framing is the real data point: the agents ran in a sandbox that could only control their own characters — no process starts, no file writes, no server commands — and the human watchers kept a dashboard with pause and stop buttons they never had reason to use. Nobody intervened on a task longer than most human workdays, four times over.
The legal risk from always-listening AI gadgets may land on the person wearing the device, not the company that sold it, according to CNBC. Amazon's new Bee Frontier wearable and Bee Eclipse card, Apple's Siri Recap on the Watch Series 12 and Meta's camera glasses all listen continuously, interpret speech and produce summaries — without storing the underlying audio. That's the twist: many state wiretap laws define interception as acquiring the contents of a private communication, not making a recording, and privacy attorney Colin Zick says a device that hears speech and writes a summary could fall inside Massachusetts-style language even with no audio file kept. A University of New Hampshire law professor, Daniel Pi, told CNBC he believes New Hampshire's statute — which covers the "aural or other acquisition" of a communication's contents — would already reach the Apple Watch's new features, and that a wearer could technically face civil or criminal liability for using one during a private conversation without everyone's consent. Prosecution risk is low, both men say; the mismatch between laws written for bugging gear and features people switch on deliberately is not. We covered the watch half of this question in September — Apple's always-listening watch runs into eavesdropping laws — and CNBC's reporting extends it to the bystander: a former CIA officer quoted in the piece argues the shift is that people can be swept into AI-powered surveillance without owning or knowingly interacting with the device doing the collecting.
What to watch: whether the Minecraft stack repeats the feat with perception stripped — raw pixels, Hard difficulty, several runs instead of one — and whether a state prosecutor or plaintiff tests the summary-is-interception theory first.
Is a $520 autonomous Minecraft win the right yardstick for agent readiness, or does perfect machine perception make it a different game entirely? Write to us via our contact page.
