Google DeepMind: four Gemini 4 Argon agents beat Minecraft from a random spawn in 27 hours, writing 2,783 programs themselves with only standard Mineflayer; a single run on Easy

Why it matters

【Key points】

Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu published "A team of Gemini 4 Argon agents beats Minecraft" (arXiv 2610.22319, submitted Oct 15); the Google DeepMind account promoted it on Oct 16 at 16:11 UTC. Four Gemini 4 Argon coding agents shared a random Minecraft survival world (Java 1.20.4, Easy, cheats off) with one goal: defeat the Ender Dragon.

They never saw pixels. Each agent wrote its own JavaScript through the open-source Mineflayer library and coordinated via in-game chat and shared memory, with no assigned roles, game logic or skill library. The dragon fell after 27h14m, 2,783 programs, 229 messages and 605 deaths. Roles emerged on their own; one agent triangulated the stronghold from two Eye of Ender throws to within 43 m, and teammates recovered each other's items. A fifth Observer agent watched for cheating and found none. At API list prices the run cost about $520 up to the win ($1,040 at standard price).

【Verifiable】

Paper: https://arxiv.org/abs/2610.22319

Google DeepMind post: https://x.com/GoogleDeepMind/status/2111127991401247033

Author post: https://x.com/maxencefaldor/status/2111023397106770026

【Our read】

1. A clean demo of long-horizon, multi-agent, self-coded work: one goal held for over a day across thousands of dependent steps, the skills real software tasks need.

2. The authors state the limits: one run (their first try), so no reliability estimate; Easy difficulty; Mineflayer hands agents exact block, mob and dragon state, far easier than pixels; and the model already knows Minecraft well.

3. Read "first" carefully: community projects have killed the dragon before, including one Claude Opus 5 agent with 37 high-level tools on Normal. DeepMind's difference is standard Mineflayer only, no human-written game tools, a random seed and four cooperating agents.