Google DeepMind AI Agents Beat Minecraft… After 27 Hours And 605 ‘Deaths’

Insider Brief

• Four Google DeepMind AI agents defeated Minecraft’s Ender Dragon after 27 hours and 14 minutes. They wrote their own code and worked together without human help during the game.
• The agents split up tasks, shared supplies and fixed mistakes. They created 2,783 programs and died 605 times.
• The successful attempt used Easy difficulty and software that gave the agents detailed information about the game. This leaves questions about how reliable they are and how well they would perform on harder tasks.

Four Google DeepMind AI agents defeated Minecraft’s Ender Dragon by writing their own programs and working together for more than a day. The AI services they used cost up to $1,040, and no human stepped in to help them play.

This result does not mean AI agents have mastered every challenge. The study suggests that agents still need to become more reliable before they can handle practical tasks that take a long time to complete.

According to the study, published on the preprint server arXiv, the team finished the task in 27 hours and 14 minutes. It created 2,783 programs, split up the work, shared supplies and recovered from repeated failures.

The experiment tested whether general-purpose AI coding agents could follow a plan through thousands of connected actions. The researchers gave four copies of Gemini 4 Argon control of separate players in the same world. Their goal was to defeat the dragon. The researchers did not give them a game strategy or assign them specific jobs.

The agents won, but they were not very efficient. Their characters died 605 times, 22% of their programs ended with an error, and the final stage took almost 14 hours. The experiment also used Easy difficulty and software that provided detailed information about the game world.

These limits affect what we can learn from the result. The paper describes one successful attempt, which the researchers say was their first launch. It does not show how often the system would succeed or prove that it could perform just as well at business tasks or software development.

Even so, the experiment gives a detailed record of AI systems organizing work, creating tools and fixing mistakes over a long task. The researchers say these abilities are relevant to complex software projects and business work, where success depends on completing many steps in the correct order.

Playing Through Code

The study used Minecraft Java Edition 1.20.4 in survival mode. Cheat commands were turned off, and the game created a random world.

Defeating the Ender Dragon requires several steps. Players collect materials, make tools, enter another region called the Nether, gather ingredients and find an underground stronghold. The stronghold contains a portal to the dragon’s realm.

The agents completed these tasks using Mineflayer, an open-source software library that lets programs control Minecraft characters. Instead of looking at the screen and pressing keys, each agent wrote short JavaScript programs to make its character walk, mine, craft items or attack enemies.

The software returned text describing the player’s location, health, inventory and actions. The agent used this information to write its next program.

Mineflayer provided basic controls and a plugin that planned paths around obstacles. According to the study, the agents chose which goals to pursue and wrote the code needed to reach them.

Their general instructions encouraged them to check that actions had worked, create tools they could reuse and save useful lessons in shared memory. The researchers also explained how to use the controls and communicate with teammates. They did not provide a game strategy.

A fifth AI agent watched the session and compared reported progress with recorded events. It did not play the game. Researchers could watch a live dashboard and pause or stop the experiment, but they reported that they did not need to intervene.

A Team Forms

The agents started discussing who should do which tasks within minutes. About 10 minutes into the game, one agent suggested splitting up the work. Another agent suggested a different plan.

As the game continued, they settled into different roles. One focused on food and supplies, another made tools and searched for important structures, and the others worked on portals, shelters and preparations for fighting.

The team exchanged 229 messages before winning. Shared storage chests let the players pool their supplies. Their chat messages and notes also saved locations, warnings and lessons about their tools.

One early discovery shows how a fix made by one agent helped the whole team. Nineteen minutes into the game, an agent warned the others about a problem with how Mineflayer handled crafting. The researchers reported that a helper function designed to fix this problem later appeared in 655 programs written by all four agents.

The agents also wrote software to keep their characters afloat, count items and stop old instructions from continuing after a program ran past its time limit. This last feature appeared in 1,519 programs.

These tools helped the team deal with a common problem in long AI tasks: an agent needs to keep useful information while switching between jobs and shortening its growing conversation history. Shared notes gave the agents a record they could use even after these changes.

Mistakes also led the agents to change their behavior. For example, after an enemy's fireball destroyed supplies that had been dropped on the ground, the agents warned one another and worked out a safer way to return.

According to the paper, the agents could also spot some incorrect reports. One agent corrected a message saying that supplies had been delivered after finding that its inventory was empty. Another questioned a teammate's claim that it had returned with materials because the server had recorded that teammate's death only a short time earlier.

These examples show that the agents sometimes checked information and corrected mistakes. They do not prove that the agents caught every incorrect claim.

Progress… But Also Heavy Losses

The team built a portal to the Nether after one hour and 42 minutes. Collecting and carrying the materials needed for the next stages turned out to be harder.

It took more than six hours to bring one important ingredient back to the starting area. Deaths and lost supplies repeatedly slowed the team down.

To find the stronghold, one agent tracked the paths of two Eyes of Ender, items in the game that point toward the structure. It measured their directions from two different locations and worked out where the paths crossed. According to the paper, its estimate was 43 meters away from the actual portal room.

The team entered the dragon’s realm 13 hours and 33 minutes after the run began. Finishing the task took almost as long again.

Before they could defeat the dragon, the agents had to destroy 10 crystals that restored its health. One player destroyed six crystals, another destroyed three, and a third destroyed the last one. The dragon was defeated about 40 minutes after the final crystal was destroyed.

Most of the deaths during the run were caused by monsters. Others happened because of drowning, falls, lava, fire and the dragon itself. The server also restarted six times, stopping the game for about 20 minutes in total.

The experiment also required a lot of computing power and money. Before winning, the four players made about 24,000 model calls and generated around 16 million output tokens, which are units used to measure how much text an AI system processes.

The researchers estimated that the model usage cost about $520 at introductory public API prices, or $1,040 at standard prices. These are estimates based on listed prices, not the actual cost of running the model.

Remaining Questions

The software interface gave the agents important advantages over human players and AI systems that have to play by looking at images.

Mineflayer gave the agents exact information about nearby blocks and creatures, including blocks hidden behind solid rock and the dragon’s internal state. Its navigation software also planned paths around obstacles. The researchers acknowledge that playing through a screen with a keyboard and mouse is much harder.

The model also had some knowledge before the experiment began. Gemini 4 Argon had learned about Minecraft and Mineflayer during training. The experiment tested whether it could use that knowledge to write working programs and coordinate actions over a long period.

The researchers also recognize that earlier community projects had reported defeating the dragon with other AI setups. They say their approach was different because of the tools and conditions they used, including the standard Mineflayer library, a random world and no extra game-specific strategies or skills.

More tests are needed to find out whether the system can succeed in different worlds and across repeated attempts. Comparing the team with a single agent would help show how much working together contributed to the result. Harder difficulty settings or controls based on what the agents see on screen would test them under more challenging conditions.

The DeepMind research team included Maxence Faldor, Satinder Singh and Roberta Raileanu.

For more technical details, read the paper on arXiv. It is important to remember that arXiv is a preprint server, where researchers can share their work and get feedback quickly. Papers posted there have not necessarily passed official peer review. Peer review is an important part of science because it helps check whether research findings are reliable.