Four Gemini 4 Agents Beat Minecraft's Ender Dragon in 27 Hours After 605 Deaths
量子位 · wechat · 2026-10-24
Google DeepMind dropped four Gemini 4 Argon agents into a random Minecraft map with a single instruction: defeat the Ender Dragon—no task assignments, no walkthroughs. Reading game state via Mineflayer and writing their own JavaScript, the agents chatted 229 messages, wrote 2,700+ lines of code, and won after 27 hours and 605 deaths—including the pivotal debugging moment when one agent discovered end crystals were healing the dragon and piled dirt blocks to climb up and punch all 10 crystals out by hand. A netizen calculated the run cost $520: 1.35B input tokens, 16M output tokens, 24k model calls (91% of inputs cache-served). The experiment probes long-horizon agency: decomposing a vague goal into thousands of actions, remembering it across days, and recovering from teammate deaths and tool failures. Paper: arxiv.org/abs/2610.22319.
More from coding & agent
- Claude Code audit finds 214,856 "permanently deleted" photos lurking in Android hidden cache — Loo_Atreides · 2026-10-26
- Someone sent the first customer fax via ChatGPT over Onetime Fax's new MCP server — tdhopper · 2026-10-26
- China Telecom's XingChen lab open-sources Xing4.0-29B-Intent, a 29B MoE with 4B activated params — FellMentKE · 2026-10-26
- Codebases sized to fit one context window dramatically improve agent outcomes — julianharris · 2026-10-26
- 42-Minute Video Breaks Down Why AI Agents Fail: Mostly Not the Model's Fault — JohnSavill · 2026-10-26
- Giraffe: Open-Source Baseline Testing for OpenAI-Compatible LLM Endpoints — ImmediateBig8479 · 2026-10-26