Four AI agents beat Minecraft in 27 hours, DeepMind researchers say

Four AI agents kept a shared Minecraft mission going for more than a day, dividing work and writing their own programs until they defeated the Ender Dragon. Google DeepMind researchers say the Gemini 4 Argon team finished in 27 hours and 14 minutes, without human input during gameplay or a researcher-assigned game plan.

The paper, dated October 15, 2026, is by DeepMind’s Maxence Faldor, Satinder Singh and Roberta Raileanu. They set the team one objective: defeat the dragon from a random starting world. The experiment tested whether general-purpose coding agents could maintain a plan across thousands of dependent actions rather than follow human-written game logic.

Programs, not pixels

Each agent controlled one player through Mineflayer, an open-source JavaScript library for Minecraft bots. Instead of watching a screen or using a keyboard, it wrote a short program to walk, mine, craft or fight. The program returned text describing its position, health, inventory, discoveries or errors. The agent then chose what to write next.

Mineflayer supplied basic actions, with a pathfinding plugin for navigation, but did not decide which actions would win the game. Programs ran until their task finished or an agent-selected time limit expired. The median program lasted 33.5 seconds; each player had code running about 40% of the time, with the rest spent reading, thinking and writing.

The researchers’ setup used Minecraft Java Edition 1.20.4 in survival mode on Easy difficulty, with cheat commands disabled. Each agent also had ordinary software-work tools, including a shell, file editing and version control. Its general instructions required checking results, testing new tools on short tasks and preserving lessons in memory.

A team that assigned its own jobs

Coordination began early. Ten minutes after the first program, an agent called janesick proposed a division of labor in the game’s chat. Within a minute, auratian offered a different plan and gyldenia acknowledged it. Roles then settled into concrete jobs:

• Gyldenia scouted for food, beds and flint.
• Bulicame went underground to find diamonds.
• Janesick smelted iron and made tools.
• Auratian prepared the portal to the Nether.

Chat carried messages to every player, while shared notes preserved coordinates and lessons. The agents also recovered one another’s items after deaths. Their memory instructions addressed a separate continuity problem: when a conversation grew too long and was compressed, work in progress and next steps needed to survive.

2,783Programs

JavaScript programs written and run before victory.

229Chat messages

Messages exchanged among teammates during the run.

605Deaths

Player deaths along the way to defeating the dragon.

A win with boundaries

Progress was uneven. The team entered the End, the dragon’s dimension, after 13 hours and 33 minutes. It did not destroy the last of the ten crystals that heal the dragon until 26 hours and 34 minutes. The final victory followed 40 minutes later. Total elapsed time included six server restarts that paused the world for about 20 minutes.

A fifth agent, the Observer, checked player claims against the record and produced a live report without playing. It was instructed to flag fabricated results or attempts to exploit the evaluation. The authors say it flagged no such exploitation, and that its findings and their manual log reviews identified no safety incidents. Researchers could pause or stop the run.

The authors call this, to their knowledge, the first end-to-end autonomous Minecraft completion using standard Mineflayer without pre-scouted worlds or human-written, game-specific tools. They connect the planning and coordination demands to software engineering and enterprise work. But the demonstrated result remains a sandboxed game completion, not evidence that the team can perform those real-world tasks reliably.