Four Gemini 4 Argon agents finished Minecraft on their own: 27 hours 14 minutes, 605 deaths and 2,783 self-written programs

Pin the source first. The paper is titled 'A team of Gemini 4 Argon agents beats Minecraft', by Maxence Faldor, Satinder Singh and Roberta Raileanu, arXiv 2610.22319, cs.AI, submitted October 15, 2026. TechRepublic and Tom's Hardware covered it on October 20, and most numbers circulating in Chinese-language coverage trace back to those two write-ups of the paper. Every figure here comes from the abstract or those reports; anything the paper does not state is labelled as a media-report figure.

How they played: no pixels, no simulated keystrokes. Each agent controlled a player through Mineflayer, an open-source JavaScript library, by writing short code snippets to walk, mine or fight, then reading text feedback after each action before deciding the next step. The interface is code and text, not images. That detail matters, because every later argument about transferring this to real work runs into it.

The hard numbers: four player-side agents plus a fifth bot dubbed the Observer that watched the session to confirm no rules were broken. Across the run they wrote 2,783 programs, exchanged 229 in-game chat messages, died 605 times, took 27 hours and 14 minutes, and saw 22% of their programs throw errors at execution. The start was a randomly generated survival world on Easy difficulty, with no pre-scouted seeds and no game-specific tooling written by humans.

Roles emerged rather than being assigned. As reported, Janesick manufactured tools, gyldenia gathered food, auratian activated the portal, and bulicame built defences and landed the final blow. The clearest example of coordination is a failure: auratian drowned while carrying all 12 Eyes of Ender, and teammates calculated the despawn countdown on the dropped items and rushed over to recover the gear. None of that was scripted; they worked it out.

Cost: per the reports, the run consumed about 1.35 billion input tokens, roughly $520 at introductory list rates, or about $1,620 for the full session at standard pricing. That figure is worth remembering because it sets the current order of magnitude for long-horizon agents: one successful multi-dozen-hour run, a four-figure bill. Anyone planning to put such agents into production has to model that line item first.

Why this matters more than 'AI plays a game'. The paper's own framing is software engineering and enterprise work: real coding and business tasks also require holding one coherent plan across thousands of dependent steps, and also require multiple agents to divide work, share what they learn and recover each other's mistakes. Minecraft here is a controllable testbed, not the point. What it measures is whether a plan can survive a day.

The limits have to be stated plainly. TechRepublic's first point is the nature of the environment: block coordinates and entity health were exposed directly through an API in this sandbox, so the agents did not have to perceive a messy graphical interface the way a human would. The second is the 22% program error rate: in a game you can retry, while in enterprise workflows touching money, sensitive data or production systems there is no such margin. The third is that humans watched a live dashboard during the run; they did not intervene, but that is not the same as unsupervised.

Boundaries, and what this article does not claim. First, 'first' is the authors' own characterization — the wording is 'to our knowledge', not a third-party verified record. Second, the paper was submitted October 15 and drew wide media coverage on October 20; this article treats those as two separate events. Third, nothing here supports the conclusion that AI agents can replace human engineers; the paper says real software engineering demands the same capabilities, and needing them is not the same as having them. Fourth, this article does not claim the full code and configuration are published to the point where an outsider could reproduce the run in one click.