Google DeepMind's Gemini 4 Argon Beats Minecraft Autonomously in 27 Hours
Four general-purpose Gemini coding agents autonomously killed the Ender Dragon in 27 hours, writing 2,783 programs and dying 605 times with zero human intervention.
- Four Gemini 4 Argon agents autonomously defeated the Ender Dragon in 27 hours and 14 minutes.
- They wrote 2,783 JavaScript programs and died 605 times with zero human intervention.
- Agents coordinated via in-game chat and shared memory, spontaneously dividing roles within ten minutes.
- Interface was Mineflayer, with no Minecraft-specific tools or pre-scouted seeds provided.
- Team built 42 reusable helper functions on the fly, bootstrapping their own Minecraft SDK.
- One agent triangulated the hidden stronghold by intersecting two Eye of Ender trajectories.
Four coding agents beat Minecraft after 27 hours
Four general-purpose coding agents entered a randomly generated Minecraft world with one instruction: kill the Ender Dragon. They finished in 27 hours and 14 minutes without human strategic guidance, a pre-scouted seed, or human-written Minecraft-specific tools.
Google DeepMind researchers Maxence Faldor, Satinder Singh, and Roberta Raileanu describe the experiment in a research paper. To their knowledge, it is the first autonomous, end-to-end Minecraft completion by AI agents under those conditions.
| Measure | Result |
|---|---|
| Active players | 4 |
| Observer agents | 1 |
| Completion time | 27 hours, 14 minutes |
| Generated programs | 2,783 |
| Deaths | 605 |
| Chat messages | 229 |
| Reusable helper functions | 42 |
| Difficulty | Easy |
Minecraft stretches plans across thousands of actions
Completing Minecraft from a fresh world requires a long chain of dependent objectives. Players must gather resources, craft equipment, enter the Nether, obtain blaze rods and Ender pearls, create Eyes of Ender, locate a stronghold, activate its portal, survive the End dimension, destroy healing crystals, and kill the dragon. Death can scatter or destroy inventory needed for later stages.
The paper compares the agents’ 27-hour run with a 45-hour median for Minecraft’s main story on HowLongToBeat and a random-seed speedrun record under seven minutes. The agents operated without prior knowledge of the world and finished faster than the cited median, with far more failures and execution overhead than expert human play.
Earlier Minecraft agents, including NVIDIA’s Voyager, used Minecraft-specific primitives and curriculum mechanisms. This experiment tested whether a general coding-agent system could construct its own tools and preserve a plan across thousands of generated scripts.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
