Four Gemini 4 Argon agents finish Minecraft solo, killing the Ender Dragon with no human input
A quartet of Gemini 4 Argon agents cleared Minecraft autonomously, writing 2,783 programs and dying 605 times across a $1,620 session.
A team of four AI agents powered by Google's Gemini 4 Argon has completed Minecraft end to end without a person touching the controls, according to The Decoder. The agents killed the Ender Dragon, the game's final boss, in a run where no human assigned roles or handed down a strategy. It is the first reported fully autonomous clear of the title by a multi-agent setup.
The mechanics of the run are as telling as the result. The agents wrote 2,783 programs between them, swapped 229 chat messages to coordinate, and died 605 times before finishing. That death count works out to roughly one respawn every few minutes of play, a reminder that the agents learned the game by failing at it repeatedly rather than following a scripted plan.
What separates this from earlier Minecraft agents is the absence of scaffolding. Prior systems leaned on hand-built task hierarchies or human-curated skill libraries to break the game into manageable steps. Here the four agents negotiated their own division of labor through conversation, deciding who would gather resources, who would craft, and who would push toward the End. The 229 messages are the paper trail of that negotiation.
The run also came with a price tag: $1,620 in model usage for a single completion. That figure puts the cost of autonomous long-horizon play in plain terms. Minecraft is cheap to simulate but expensive to reason about, and a four-figure bill for one dragon kill shows how much inference the task consumes when agents must plan, retry, and recover from hundreds of deaths.
Minecraft has become a standard proving ground for agentic AI because it rewards exactly what current models struggle with: long-term planning, tool use, memory across sessions, and cooperation under uncertainty. A clear with no predefined roles suggests Gemini 4 Argon can hold a shared goal together long enough to matter. It also raises the question of how much of the result came from the model versus the harness built around it.
The next thing to watch is replication. A single run at this cost is an anecdote until other teams reproduce it on different models and report their own token bills and death counts. If autonomous multi-agent clears become routine, the same coordination methods will be tested on tasks with real stakes, from software engineering to warehouse robotics, where a 605-death learning curve is not an option.