Four Google AI Agents Beat Minecraft With No Human Help

Four Google AI agents just beat Minecraft with no human help, and the numbers behind the run are the real story. Placed in a fresh survival world with one goal, defeat the Ender Dragon, a team of Gemini 4 Argon agents pulled it off in 27 hours and 14 minutes, writing and debugging 2,783 of their own JavaScript programs, dying 605 times, and coordinating everything through in-game chat and a shared memory log. No human assigned roles. No human wrote game logic. It is the first end-to-end autonomous completion of Minecraft, published by Google DeepMind researchers on arXiv on October 15, 2026.

Why our team cares: this is the clearest public demonstration yet of what “agentic AI” can actually do over long, messy, multi-day tasks, not just answer questions. Here is how the run worked and how Canadians can try the model behind it.

What is Gemini 4 Argon?

Gemini 4 Argon is Google’s frontier reasoning and execution model, the member of the Gemini family built to carry long, complicated workflows from start to finish rather than answer one prompt at a time. It powers multi-agent software development in Google Antigravity, agentic coding in AI Studio and Android Studio, and advanced reasoning in the Gemini app, where it is included with a Google AI Pro subscription.

Minecraft was chosen as a stress test for exactly those traits. Beating the Ender Dragon takes thousands of interdependent steps: punching trees, crafting tools, mining diamonds, building a fortress in the Nether, and finally a boss fight. One wrong decision hours earlier can sink the whole run. That is the shape of real-world work, which is why the gaming framing undersells it.

How the run actually worked

The detail that gets our team: the agents learned from each other’s deaths. When one fell, the others logged what killed it and adjusted. Recovering a teammate’s dropped gear is the kind of emergent teamwork nobody programmed in, and it is the behaviour that matters for real agent systems.

One more number for perspective: coverage of the paper put the compute cost around US$1,620 at announced Gemini 4 Argon pricing. Autonomy is real, and it is not free.

Where the internet landed on it

Reaction split the way it usually does. The milestone framed as a genuine step for multi-agent planning and code generation across thousands of steps. The skeptic corner called Gemini 4 Argon “slow but thoughtful”, pointing out that 27 hours and 605 deaths is not exactly a speedrun. Both are true, and both miss the point: the run was never about beating a game quickly. It was about proving a model can hold a goal, recover from hundreds of failures, and organise a team without a supervisor for more than a day of continuous work.

Can you try Gemini 4 Argon in Canada?

Yes. Gemini 4 Argon is included with a Google AI Pro subscription at $26.99 per month in Canada, with much higher usage limits on Google AI Ultra. You can use it in the Gemini app for advanced reasoning and in Google Antigravity for agentic coding. Canadians should note a few features on Google’s plans page are marked US-only, though Argon access itself is not one of them. The full tier-by-tier breakdown, including which features skip Canada, is in our guide to Google AI Ultra in Canada.

And Argon is not just playing games. It is the planning brain that feeds task plans to Google Genie 4, DeepMind’s new world model, and the backbone of agentic coding across Google’s developer tools. The Minecraft paper and that product integration are the same story told twice: models that plan, act and recover.

FAQ