Press Esc to close

Gemini 4 Argon agents beat Minecraft on their own, dying 605 times

Gemini 4 Argon agents beat Minecraft on their own, dying 605 times
The End portal in the Minecraft stronghold with all twelve Eyes of Ender in place, rendered by Google DeepMind from the Gemini 4 Argon agents' run

Image: Google DeepMind

Four AI agents just beat Minecraft, and nobody told them how. Google DeepMind says a team of four Gemini 4 Argon agents started with bare hands in a randomly generated world and killed the Ender Dragon, the game's final boss, 27 hours and 14 minutes later.

The agents never looked at the screen or pressed a key. According to the technical report by DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu, each agent played by writing short programs for Mineflayer, an open-source bot library, then reading back what happened: its position, health, inventory or an error message.

DeepMind says these were its everyday general-purpose coding agents, not a system built for Minecraft. Nobody assigned roles or wrote game logic. Ten minutes in, the agents had split the jobs over chat: one made tools, one gathered food, one dug for diamonds and one built a portal to the Nether, the game's fiery second dimension.

Google DeepMind chart of the four Gemini 4 Argon agents and the 13 steps of their Minecraft run, from bare hands to the Ender Dragon at 27 hours 14 minutes, with 605 respawns

Image: Google DeepMind (icons: Pixel Perfection texture pack, CC BY-SA 4.0)

Some moments read like a clever human player. To find the hidden stronghold that leads to the final dimension, one agent threw two Eyes of Ender from different spots, tracked each flight in code and calculated where the paths crossed. Faldor writes that the estimate landed 43 meters from the target, 1.3 kilometers away.

Google DeepMind diagram showing how agent janesick triangulated the stronghold from two Eye of Ender throws, landing 43 meters from the portal room

Image: Google DeepMind (icons: Pixel Perfection texture pack, CC BY-SA 4.0)

The dragon fight was the hard part. It took nearly 14 hours, and the agents died to the dragon 69 times, writing down lessons between attempts. One eventually worked out that crystals on top of towers were healing the dragon, and that it could reach them by stacking dirt instead of digging. The final blow came from a player down to 2.5 of 20 health points.

It was far from a clean run. The team died 605 times, mostly to monsters, and 22% of its 2,783 programs ended in an error. One agent even retracted a success its own program had wrongly announced: "My previous chat was an unguarded script branch."

Google DeepMind charts of what the agents' 229 chat messages carried and the causes of their 605 respawns, led by monsters at 351

Image: Google DeepMind

For scale, the report notes that the median human takes about 45 hours to finish Minecraft's main story, per HowLongToBeat, while the random-seed speedrun record is under seven minutes.

The bill was modest too. DeepMind estimates the four players used about 1.35 billion input tokens and 16 million output tokens, roughly $520 at Gemini 4 Argon's introductory API price, or $1,040 at the standard rate. That works out to under 20 cents per program, though it's a list-price estimate, not Google's real serving cost.

Earlier AI dragon kills leaned on big shortcuts. The paper cites a community run where a single Claude Opus 5 agent used 37 hand-built Minecraft tools, and another on Peaceful mode with a fixed, pre-scouted world. DeepMind says its agents got only the standard Mineflayer library and its pathfinding plugin.

That still isn't how people play. Mineflayer tells the agents exactly which blocks and creatures are nearby, even behind solid rock, and the plugin plans their routes. They played on Easy, and Argon already knew plenty about Minecraft from its training. It was also a single run, the first one tried, so it says nothing about reliability, and nobody outside Google has reproduced it yet. The researchers say a fifth "Observer" agent and a manual review of sampled logs found no cheating.

So why does Google care about a video game? DeepMind argues that sticking to a plan for over a day, recovering from failure and splitting work across a team are the same skills long software and research projects need. It plans to use the setup for coding, science, engineering and robotics, and next wants agents to play from pixels, as people do.

Comments

Reader support

No ads. Just independent reporting.

Techcity Authority runs without advertising. Hosting, tools, and the hours behind each piece are supported by readers instead. Even a small contribution goes a long way — and every supporter is credited by name below for the month they give.

Support the site

Monthly via PayPal · cancel anytime

Secure checkout on PayPal. Your name appears here for the month you support.

This month’s supporters

    No supporters yet this month — yours could be the first name here.