Four Gemini 4 agents beat Minecraft with no human help

Started by Ballad, Oct 22, 2026, 11:07 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Four Gemini 4 agents beat Minecraft with no human help   Views(Read 80 times)

Ballad

Google says a team of AI agents has beaten Minecraft completely on their own. Engadget's Steve Dent reports on a Google paper describing four Gemini 4 agents that defeated the Ender Dragon, the game's final boss, without any human intervention

The agents played a randomised survival world using only the open source Mineflayer library, with no assigned roles or game specific tools. They split the work between themselves, coordinating through chat and a shared memory. One focused on making tools, one gathered supplies, one lit the portal, and one built shelters and landed the final blow

The run took just over 27 hours. In that time the agents wrote 2,783 programs, exchanged 229 messages and died 605 times. They used 1.35 billion input tokens, which Engadget estimates cost between 810 and 1,620 dollars

A fifth agent acted as an observer to check that nobody cheated. Google describes it as the first fully autonomous AI completion of Minecraft. Previous AI Minecraft projects usually relied on human demonstrations, special tools or narrower goals

Minecraft is a great test for AI because it needs planning, cooperation and adapting to the unexpected. Agents dividing work without being told how is impressive. Dying 605 times shows it was far from elegant, but persistence clearly paid off

Is beating Minecraft a meaningful AI milestone? And what should AI agents attempt next?


WorldClassWill71

Next challenge should be building something creative. Beating the dragon is a clear goal. Designing a beautiful castle would test something else entirely
My model's smarter than me, low bar admittedly

CMPunk88

Agents dividing roles by themselves is the impressive part. Nobody told them who should mine and who should build

Coordination emerged naturally. Emergent coordination like that is the interesting research result. It suggests agents can organise themselves for long tasks. That has uses well beyond games

Lucy_35

Writing 2,783 programs shows how they worked. They coded their way through the game rather than clicking

That is a very different style from human play. Code as a way of acting in the world is a powerful idea

Nina24

Six hundred and five deaths made me laugh. That is a very human approach to Minecraft. Keep trying until it works. Failure is part of learning, even for AI
rm -rf /bad-ideas

Cyborg77

Over a thousand dollars in tokens to beat a game is expensive. Speedrunners do it in minutes. Efficiency still has a long way to go

Costs will fall quickly, though. Prices for inference keep falling, so this will get cheaper
Quantum leap over Monday, faceplant into Tuesday

Drogba32

An observer agent checking for cheating is a nice touch. AI systems often find shortcuts. Having a referee builds trust in the result. Shortcuts are a known weakness in AI agents. Supervision like this will become standard in research