Gemini 4 Argon, week two: four agents beat Minecraft from scratch in 27 hours with no human help

A week after Sundar Pichai announced Gemini 4 Argon as Google's frontier model for long-horizon work, DeepMind has published a showcase result. Four autonomous Argon agents played Minecraft from a fresh world to the Ender Dragon kill in 27 hours, with no human intervention. The agents divided labor, warned each other about threats, and retrieved items from teammates who had died. Those are coordination behaviors that are hard to script.

The key distinction DeepMind draws is generality. AlphaGo and AlphaStar were purpose-trained systems that dominated one game. Argon is the same general model that is sold through the API and already used for coding, cyber defense and robotics. Earlier Argon results showed it converting video and sketches into MuJoCo physics code and lifting Spot quadruped manipulation success from 56% to 75%. Minecraft is a long-horizon, partially observable environment, so finishing it means sustained planning over thousands of steps.

The result lands amid a broader agent arms race. OpenAI this week added Chrome-tab reading and task completion to ChatGPT. Meta is pitching Muse Wonder for autonomous workflows. xAI is restructuring Grok around a four-model router. Multi-agent coordination is becoming the benchmark that labs want to be seen winning, rather than single-turn question answering.

Caveats apply. Showcase runs are curated, and DeepMind has not published how many attempts preceded the successful 27-hour run or the tool scaffolding the agents used. A related DeepMind claim that more than 75% of its production code is AI-written drew pushback: community analysis noted only 11.9% of PRs are fully delegated and 88% still need human steering. Watch for reproductions and a technical report with failure rates.