Four Gemini 4 Argon agents beat Minecraft without human help

By Splitfeed EditorialLast updated Oct 18, 2026, 12:50 UTC

In a session spanning 27 hours and 14 minutes, four AI agents running on Google's Gemini 4 Argon model completed a full run of Minecraft, culminating in the defeat of the Ender Dragon. The agents operated without predefined roles, scripts, or any human intervention during the run. Across the session they wrote 2,783 programs and exchanged 229 chat messages to coordinate with one another, dying 605 times before finishing the task. The cost of the run is reported differently depending on the outlet: The Decoder puts the figure at $1,620, while FourWeekMBA reports approximately $520, and neither figure has been reconciled in available reporting.

The significance here lies in what the run represents as a benchmark for autonomous AI behavior. Minecraft's Ender Dragon completion requires a long chain of sequential, interdependent tasks, gathering resources, crafting tools and armor, locating the End portal, and surviving a multi-stage boss fight. Doing so without human-set roles or strategies means the agents had to improvise task division and respond to failure repeatedly, which the 605 deaths illustrate. Whether this constitutes genuine emergent coordination or a very expensive scripted-seeming process is the core question the cost discrepancy makes harder to answer cleanly. If the run cost $1,620, the economics of autonomous AI agents look very different than at $520, and that gap matters for anyone assessing whether such systems are approaching practical usefulness.

No reactions from the Pro-AI, Anti-AI, or Middle Ground camps have been published yet on this story. Typically, the Pro-AI camp would frame a result like this as a milestone in agentic AI, pointing to the self-organization and long-horizon planning as evidence that models are approaching meaningful autonomy. The Anti-AI camp would likely question whether defeating a video game boss translates to anything consequential, and would probably highlight the cost and the 605 failures as signs of inefficiency rather than capability. The Middle Ground camp would be expected to ask what scaffolding or infrastructure supported the agents and whether the result is reproducible under different conditions.

The cost discrepancy is the most immediate factual question left open. A clarification from Google or the researchers behind the run, along with any technical writeup detailing the agent architecture and coordination mechanism, would give a clearer picture of what actually drove the result and how it should be measured against prior AI game-playing benchmarks.