aifollow.news Search
Back IT之家 科技新闻
IT之家 科技新闻· · Original publication time

Google researchers say four Gemini 4 Argon agents autonomously completed MinecraftMachine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI-assisted summary

Google researchers say four agents coordinated through in-game chat and planned their own actions in a randomly generated world, defeating the Ender Dragon after 27 hours. They died 605 times; 1.35 billion input tokens would cost about $1,620 by ordinary user pricing.

Article · AI translation

IT Home reported on October 19 that Google researchers recently published a paper saying they had used 4 Gemini 4 Argon agents to complete Minecraft, defeating the Ender Dragon in just 27 hours.

Google placed the 4 Gemini 4 agents in a randomly generated world, where they could coordinate with one another through the in-game chat box. But the researchers gave them no strategy, stating only the final objective; the agents planned all their actions themselves. The researchers could observe them through a dashboard and issue commands to pause or terminate the test, but were not allowed to intervene.

After the game began, the agents quickly divided up different tasks: some built Enderman farms, some gathered food, and some focused on exploration. They shared combat tips and retrieved lost equipment after teammates died.

Over the course of the game, the 4 agents sent 229 messages, ran 2783 programs, and died 605 times. It took them 13.5 hours to reach the End, and they spent all the remaining time defeating the Ender Dragon. The process used 1.35 billion input tokens and, calculated at standard user rates, cost about US$1,620 (IT Home note: approximately 10,842 yuan at the current exchange rate).

The research team said: “To our knowledge, this is the first time AI agents have completed Minecraft from start to finish entirely autonomously. The entire process used only the standard Mineflayer library, with no seed map or human assistance.”

Found an error? Send a correction