Four Google AI agents beat Minecraft’s Ender Dragon after 27 hours

Google researchers say four Gemini 4 Argon AI agents defeated Minecraft’s Ender Dragon without human help, finishing a full run in just over 27 hours in what the team says may be a world first.

The test started with a random Minecraft seed and a single goal: beat the Ender Dragon. To do that, the agents still had to work through the game’s usual progression, including gathering Ender Pearls and Ender Rods, crafting Eyes of Ender, finding a Stronghold and completing the steps needed to reach The End. Google said the agents were not given a prebuilt strategy beyond the objective.

Google’s setup for the run

Each agent controlled a player through Mineflayer, an open-source library, and the team coordinated through in-game chat and shared memory. The researchers said no one assigned roles or wrote game logic. In their words:

“No one assigned roles or wrote game logic.”

They said the agents planned, wrote and debugged every program themselves.

Researchers could watch the run on a live dashboard and pause or stop it at any time, but they did not step in. A fifth agent monitored the process to make sure the others were not cheating. Google said the point was to observe how an agent team behaves when every step is visible and the task has to be completed without human direction.

The agents divided the work quickly. Some built Enderman shelters for farming materials, others gathered food, and others handled exploration. As the run continued, they shared combat tips and helped recover lost gear after deaths. The team had to keep its plans aligned across the full playthrough, even as the game kept throwing new problems at them.

What the agents had to do before the boss fight

Reaching the End was only part of the job. Before the final fight, the agents had to move through Minecraft’s normal survival and progression systems, from early tools to the farming and crafting needed for Eyes of Ender. Google said that sequence took a lot of work, because the path to the Ender Dragon is built around a chain of dependent tasks rather than a single combat encounter.

According to the paper, the agents reached The End in around 13.5 hours. The rest of the run was spent trying to kill the boss, with a large share of that time disappearing into repeated deaths and recovery. Over the full attempt, the agents exchanged 229 messages, ran 2,783 programs and died 605 times.

The numbers show how much coordination the task demanded. The agents had to gather materials, make decisions, recover from setbacks and keep moving toward the same goal over a long stretch of play. Even when they failed, they had to keep the overall plan intact long enough to make progress again.

Google also said the run used 1.35 billion input tokens. By the researchers’ estimate, that would cost a normal user around $1,620. Broken down further, the figure works out to roughly $60 per hour overall, or about $15 an hour for each of the four agents. That is a high bill for a task that still ended with a lot of dying before the dragon finally fell.

Why the researchers ran the experiment

The team said the goal was not to chase a Minecraft speedrun record. The current record is six minutes and 36 seconds, which puts the 27-hour autonomous run in a very different category. Instead, Google said it wanted to study how multiple agents coordinate in a setting where the whole process can be inspected from start to finish.

“To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents,” the team wrote in the paper.

They added that the run used only the standard Mineflayer library, with no pre-scouted seeds or other game-specific tools written by humans. That matters because the system had to cope with Minecraft’s open-ended structure on its own, rather than relying on a custom shortcut designed for the test.

The researchers also connected the game test to broader agent work. They said the skills needed to solve Minecraft — coordinating a team, recovering from failure and pursuing a goal for more than a day — are the same kinds of skills agents need to work reliably alongside people. The experiment let them see how the system handled visible, step-by-step progress when the task had to be completed end to end.

The run ended with four agents, one observer and a lot of failed attempts, but the result was still a complete defeat of the Ender Dragon without human-written game logic guiding the way. Google’s paper frames that as a test of planning, persistence and coordination under pressure, not just as a novelty inside a sandbox game.

That makes the result useful as a benchmark for how agent teams behave when a job is long, public and easy to audit. Minecraft gave the researchers a closed world with clear goals, visible failures and enough moving parts to show whether the bots could keep working together after repeated mistakes. The answer, at least in this run, was yes.