TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Decoder reports that four agents running Gemini 4 Argon completed Minecraft by defeating the Ender Dragon, without human input during the run. The reported session lasted 27 hours and 14 minutes, involved 2,783 JavaScript programs and 605 deaths, and cost $1,620 at a stated future standard price.
Four agents powered by Google’s Gemini 4 Argon reportedly beat Minecraft by defeating the Ender Dragon, completing a 27-hour, 14-minute run without human input, according to a report by The Decoder. The reported test matters as a demonstration of AI agents coordinating over an extended, open-ended task, although the available account does not provide enough detail to independently assess the run or its limits.
The report says the Google DeepMind team placed the four agents in a randomly generated world without assigning them roles or strategies and without supplying game-specific tools. The agents divided tasks, warned one another about threats, retrieved items after teammates died and recorded mistakes in shared memory, according to The Decoder’s account of the work by Faldor and co-authors.
Across the session, the agents reportedly wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times before completing the game. The report describes the outcome as the first time four such agents beat Minecraft without human input; that first-time characterization is the report’s claim and cannot be independently verified from the supplied material alone.
The Decoder puts the run’s cost at $1,620, using the standard Gemini 4 Argon price that it says takes effect in 2027. That is a projected-price basis, not necessarily the cost under any different pricing arrangement. The account does not specify the agents’ hardware, the exact evaluation protocol, or whether the full run and its completion were independently reproduced.
A Long-Horizon Test for AI Agents
Minecraft requires players to gather resources, craft tools, respond to threats and work toward a distant objective across a changing world. The reported run therefore offers a test of whether agents can plan, adapt and coordinate over time, rather than answer a single prompt or complete one tightly bounded game move.
The reported behaviors—sharing warnings, recovering equipment and keeping a record of mistakes—suggest practical forms of cooperation that could matter beyond games. But a successful game run does not establish that the agents can reliably handle real-world tasks. The result is best read as a demonstration under one reported setup, not evidence that general-purpose AI systems can operate independently in every complex environment.
The cost and duration also matter. More than 27 hours and a reported $1,620 at a future standard price indicate that this kind of sustained agent activity may consume substantial time and computing resources. The result may be technically notable, while its practicality depends on performance, repeatability and cost in other settings.
As an affiliate, we earn on qualifying purchases.
From Game Benchmarks to Agents
Minecraft has become a setting for testing AI because it combines exploration, resource management and survival with a clear end goal. The Decoder frames this result as part of a broader shift: game-playing systems that once required purpose-built AI may now be approached with general-purpose models operating as agents. That is the report’s interpretation of the trend, not a conclusion established by this single run.
The reported approach is notable for the absence of assigned roles and game-specific tools. It suggests the team tested whether agents could organize tasks themselves, rather than following a human-written plan. Still, the account does not describe the model’s instructions, software interfaces or any safeguards and constraints used. Those details would help readers understand how much of the achievement came from the model and how much from the surrounding system.
“The four agents beat Minecraft without any human input.”
— The Decoder, reporting on work by Faldor and co-authors
As an affiliate, we earn on qualifying purchases.
Questions About the Run’s Limits
The source account does not provide the underlying paper or enough methodological detail to establish how the run was verified. It is not clear whether a human prepared the environment, monitored progress, intervened in software or restarted the agents, or how the team defined “without human input.” The claim may refer to the gameplay itself rather than every stage of setup.
Other open questions include whether the run can be repeated, how much variation there is between attempts, what tools or interfaces the agents could access, and whether the reported cost includes all related computation. The source also does not explain what counts as a completed run beyond defeating the Ender Dragon. These gaps make it difficult to compare the result with other agent benchmarks or judge its reliability.
As an affiliate, we earn on qualifying purchases.
Evidence Needed to Test Reproducibility
The next useful step would be publication of the full research account, including the agent setup, instructions, tools, monitoring conditions, cost calculation and evidence that the game was completed. Repeated runs in new worlds would show whether the result was reproducible or depended on a particularly favorable sequence of events.
Until those details are available, the reported achievement remains a notable single-run demonstration rather than a broad measure of agent capability. Further testing could clarify how consistently the agents coordinate, how they recover from setbacks and whether similar performance is possible at lower cost or over different tasks.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did the agents really beat Minecraft?
The Decoder reports that the four Gemini 4 Argon agents defeated the Ender Dragon. The supplied material does not include independent verification or the full evaluation details.
Were humans involved during gameplay?
The report describes the run as completed without human input and says the team did not assign roles or strategies. It does not clarify whether people monitored the session or intervened in setup and software.
How long did the run take?
The reported session lasted 27 hours and 14 minutes. The agents reportedly wrote 2,783 JavaScript programs, exchanged 229 messages and died 605 times.
How much did the run cost?
The Decoder reports a cost of $1,620, calculated using a Gemini 4 Argon standard price it says takes effect in 2027. The supplied account does not give a full cost breakdown.
Does this prove AI can handle real-world work independently?
No. The reported run is evidence of performance in one game setup, not proof of reliable autonomy across real-world tasks. Reproducibility, system details and performance in other environments remain open questions.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
