The setup was deliberately unglamorous: a random shared survival world on Easy difficulty, with each agent controlling its own player through Mineflayer, an open-source library for Minecraft bots that reports where blocks and creatures are. Nobody assigned roles, yet within about ten minutes the bots had negotiated a division of labour in chat, splitting duties for food, diamonds, iron and the Nether portal. They still died 605 times, and roughly half the run went on the final boss.

The dragon fight itself accounted for 69 deaths, because ten regenerating crystals kept healing the boss until one agent worked out that stacked dirt pillars could reach them. Once all ten crystals were down, the dragon followed 39 minutes later. Other low points included an agent drowning while carrying Eyes of Ender, though the team noted the spot and rescued the items.

The run cost an estimated $520 in tokens at Gemini 4 Argon's introductory pricing, or more than double that at list rates, with about 1.35 billion input tokens, 91% of them served from cache. The authors acknowledge earlier community dragon kills, including a single Claude Opus 5 agent that finished in 3 hours and 56 minutes with one death and about $93 in API-equivalent costs, though it used 37 ready-made high-level tools rather than building everything from Mineflayer alone. DeepMind's point is not gaming prowess but proof that Argon can handle complex, long-horizon workflows and agent coordination for real-world tasks.