Four Gemini 4 Argon agents beat Minecraft with no human help

Four AI agents running on Google's Gemini 4 Argon have beaten Minecraft for the first time without any human help, slaying the Ender Dragon in a fully autonomous run. The agents received no predefined roles and no strategies. They coordinated on their own, wrote 2,783 programs, exchanged 229 chat messages, and died 605 times. The session cost $1,620 in compute.

The result, reported by The Decoder, marks a milestone for multi-agent AI systems: a small team of language-model-driven agents completed a long-horizon, open-ended task that normally demands planning, division of labor, tool use, and recovery from repeated failure — all without a human in the loop.

What did the agents actually do?

The four agents had to take Minecraft from a fresh start to its canonical endgame goal: defeating the Ender Dragon. Along the way they:

• Wrote 2,783 programs, using code generation as their primary tool for acting in the game world
• Exchanged 229 chat messages to coordinate among themselves
• Died 605 times and kept playing after each failure
• Defeated the Ender Dragon, completing the game's central objective, with no human intervention at any point

The absence of predefined roles matters. Earlier multi-agent demonstrations typically assign agents fixed jobs — one plans, one executes, one checks. Here, according to the report, the agents organized themselves. Nobody told one agent to mine and another to fight. The division of labor, such as it was, emerged from the agents' own interactions, including the 229 messages they sent each other.

Why does beating Minecraft matter?

Minecraft has served as a standard testbed for embodied AI research for years, precisely because it resists shortcuts. The game spans survival mechanics, resource gathering, crafting, navigation, combat, and a multi-stage endgame that requires collecting rare materials, building a portal, and then killing a dragon in a separate dimension. No single reflex or scripted routine carries an agent from spawn to victory.

A fully autonomous completion therefore demonstrates several capabilities at once:

• Long-horizon planning. The path from an empty inventory to the Ender Dragon's death involves hundreds of intermediate goals strung together.
• Program synthesis as action. Writing 2,783 programs means the agents did not act through hand-coded controllers. They generated their own code to manipulate the world, debugging and rewriting as runs failed.
• Multi-agent coordination. Four agents sharing one world must avoid working at cross-purposes. The 229 chat messages show the agents communicating to align their efforts.
• Resilience. Six hundred and five deaths and a finished run means the system recovered from failure without a human resetting strategy or offering hints.

What does the cost figure tell us?

The session consumed $1,620 worth of compute. That number carries more information than it first appears to. Beating Minecraft autonomously now costs roughly the price of a mid-range workstation — not a supercomputer allocation — using a frontier-class commercial model, Gemini 4 Argon, accessed at API-scale rates.

The cost-per-success framing matters for anyone tracking the economics of agentic AI. Tasks that once required sustained human engineering effort can now be pushed to completion by renting model time. The price of autonomous completion will shape which use cases become practical: a $1,620 run that ends in a dragon kill is a benchmark data point; the same figure applied to software maintenance or laboratory automation is a business case.

How does this fit the broader agentic AI race?

The demonstration lands in the middle of an industry-wide push toward agents that complete tasks rather than answer questions. Google, OpenAI, Anthropic, and others have spent the past two years shipping systems designed to operate software, browse the web, and write code with minimal supervision. Game environments remain a favored proving ground because they offer verifiable success conditions — either the dragon dies or it does not — at far lower stakes and cost than real-world deployments.

The Minecraft result gives Google a concrete, checkable claim in that competition: a named model, Gemini 4 Argon, a fixed team of four agents, a defined goal, full autonomy, and a clean terminal state. Verifiable completions of this kind are becoming the currency by which agentic systems are compared, supplementing static benchmarks that agents increasingly saturate.

What are the limits of the result?

The report does not claim generality beyond this run. Beating Minecraft, however demanding, is still a game with fixed rules, perfect state information available to the game engine, and no irreversible real-world consequences. The 605 deaths underscore how far the agents remained from efficient play; a human speedrunner finishes the game in minutes and dies rarely or never. Success here means eventual completion through persistence and self-generated tooling, not optimized performance.

The $1,620 bill also reflects a single successful session as described. The figure does not cover the cost of earlier failed attempts, if any, or the engineering that framed the run itself. What it does establish is an upper-bound order of magnitude: frontier-model autonomy, applied to a task of this length, is priced in the low thousands of dollars per completion.

What comes next?

The same architecture — multiple agents, code generation as action, self-directed coordination — points beyond games. Any domain that can be instrumented so agents can write and execute programs against it is, in principle, addressable by this pattern. The Minecraft run suggests the binding constraints are shifting from raw model capability to cost per completed task and the reliability of long autonomous sessions. Watch the next round of results to see whether the dragon-kill pattern transfers: same autonomy, fewer deaths, lower bill.