The report states that 4 Gemini 4 Argon agents collaborated to complete Minecraft.

AT A GLANCE

The report states that 4 Gemini 4 Argon programming agents collaborated for 27 hours 14 minutes before defeating the Ender Dragon in Minecraft.

Article

Google 4 Gemini 4 Argon programming agents beat Minecraft: 27 hours, died 605 times, at a cost of approximately 520 USD Beating Newsflash:AI,Google DeepMind enabled 4 programming agents Gemini 4 Argon to enter a randomly generated Minecraft survival world, giving them just one objective: defeat the Ender Dragon. The researchers did not preassign roles or write specialized game scripts. 27 hours 14 minutes later, 4 agents independently wrote and executed 2783 pieces of JavaScript code, died a total of 605 times, and ultimately defeated the Ender Dragon.

These 4 agents did not play by looking at a screen; instead, they used the open-source library Mineflayer to read coordinates, inventories, and the surrounding environment, then wrote code to control characters to mine, craft tools, and fight. Just 10 minutes after starting, they divided responsibilities on their own, respectively searching for food, mining diamonds, crafting tools, and building portals. While searching for a hidden stronghold, one agent used throws of an Eye of Ender to calculate a location only 43 meters from the target. When teammates died, they also notified each other of locations and recovered dropped items.

The Ender Dragon was the most difficult part. The team spent nearly 14 hours on the final stage and died 69 times. They later discovered that crystals atop the towers continually restored the dragon's health, so they used dirt to elevate themselves and destroyed 10 crystals one by one. In the entire experiment, 22%of code executions resulted in errors. The paper disclosed that, before completing the game, the model was called approximately 24000 times and processed 1350000000 of input,Token of which 91%was served from cache. Based on the promotional Gemini 4 Argon price of API, the cost is estimated at approximately 520 USD, or approximately 1040 USD at the standard price.

However, this was not the first time AI defeated the Ender Dragon. This year,September a Claude Opus 5 community project had already completed a similar task, but it relied on 37 pre-developed game tools. Google's advance this time was enabling general-purpose programming agents to plan, write code, and collaborate to complete the game independently. But the experiment also had clear limitations: the game difficulty was Easy,Mineflayer could directly provide information about blocks behind stone and the Ender Dragon's internal state, and a ready-made plugin was also available for pathfinding.

Original link https://m.theblockbeats.info/flash/372343

Key points

The report states that Google DeepMind 4 programming agents were placed into a randomly generated Minecraft survival world with the goal of defeating the Ender Dragon.Gemini 4 Argon
The researchers did not preassign roles or write specialized game scripts.
4 agents independently wrote and executed 27 hours 14 minutes' worth of 2783 pieces of JavaScript code, dying a total of 605 times.
The agents used the open-source library Mineflayer to read coordinates, inventories, and the surrounding environment, and wrote code to control character actions.
The team spent nearly 14 hours on the Ender Dragon stage, died 69 times, and destroyed 10 crystals one by one.
In the entire experiment, 22%of code executions resulted in errors.
Before completing the game, the model was called approximately 24000 times and processed 1350000000 of input,Token of which 91%was served from cache.
The report estimates the cost at approximately API based on the promotional price of 520 USD, and approximately at the standard price 1040 USD。
The experiment used Easy game difficulty, and Mineflayer could provide information on some hidden blocks and the Ender Dragon's internal state; pathfinding used a ready-made plugin.
The report states that this year,September a Claude Opus 5 community project had already completed a similar task, but it relied on 37 pre-developed game tools.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

These agents did not operate by watching the screen; instead, they read structured information in the game and then generated and executed JavaScript code. They independently divide responsibilities, gather resources, craft tools, find targets, and collaborate in combat, but the process involved many deaths and code execution errors.

Why it matters to readers

This experiment demonstrates how general-purpose programming agents can plan, write code, and collaborate in a game environment; however, the environment provided additional state information and ready-made pathfinding capabilities, so this completion alone cannot establish their capabilities in other scenarios.

Beginners can understand this as 4 code-writing AI characters cooperating to complete a game task. The focus is not only on ultimately completing the game, but also on the time required, error rate, model call volume, cost, and how much assistance the experimental environment provided.

Risks and unknowns

The provided materials do not include the original paper, experimental code, or Google DeepMind formal documentation, so the experimental results cannot yet be independently verified.
Gemini 4 Argon No primary source in the materials confirms the official product information, availability, or pricing of.
The cost is estimated based on the reported API listed price, rather than explicitly disclosed actual expenditure.
The experiment used Easy difficulty, additional environmental information, and an off-the-shelf pathfinding plugin; whether the results can be generalized to other tasks remains unclear.

The report states that 4 Gemini 4 Argon agents collaborated to complete Minecraft.

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center