Four AI Agents Beat Minecraft After 605 Deaths, Spending $520 in Tokens

AI Agents Achieve First Autonomous Completion of Minecraft

Google DeepMind researchers have demonstrated that four Gemini 4 Argon AI agents successfully beat Minecraft's final boss, the Ender Dragon, in 27 hours and 14 minutes without human intervention1

1

2

2

. The agents started in a randomly generated survival world on Easy difficulty with a single objective: defeat the dragon. No roles, strategies, or game-specific tools were assigned by humans. According to the paper published on arXiv, this marks the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library without pre-scouted seeds4

4

.

The achievement showcases how agentic AI models can tackle complex, long-horizon tasks requiring sustained coordination. Throughout the run, the four agents wrote 2,783 JavaScript programs, exchanged 229 messages through in-game chat, and died 605 times, mostly to monsters2

2

5

5

. The session processed approximately 1.35 billion input tokens, with 91% served from cache, resulting in a token cost of around $520 at Gemini 4 Argon's introductory pricing, or $810 including an Observer agent that monitored for cheating1

1

4

4

.

Emergent Teamwork and Multi-Agent Coordination

Within approximately 10 minutes of starting, the AI agents exhibited emergent teamwork by proposing and agreeing on a division of labor through chat1

1

. The agent janesick focused on iron and tools, gyldenia gathered food and supplies, auratian lit the Nether portal, and bulicame handled diamonds and ultimately landed the final blow on the dragon2

2

. The agents warned each other about threats, shared coordinates, provided tips, and even recovered items from dead teammates2

2

.

This multi-agent coordination mirrors practices familiar from effective human teams, despite nothing in the agents' instructions mentioning specific roles3

3

. When setbacks occurred, such as a Ghast fireball destroying dropped items, the team diagnosed the cause and wrote rules to prevent future occurrences1

1

. The researchers emphasize that these capabilities for long-term goal adherence and coordination are essential for complex real-world software engineering workflows and enterprise tasks3

3

.

JavaScript Programs Replace Traditional Gameplay

Rather than controlling characters directly through pixels and keystrokes, each agent operated by writing short JavaScript programs using the open-source Mineflayer JavaScript library for Minecraft bots3

3

4

4

. The agents received text-based feedback containing status data about their position, health, and inventory rather than visual information1

1

. This approach demonstrates the agents' coding capabilities rather than traditional game-playing skills.

Of the 2,783 programs written, 42 helper functions were kept alive between programs, while 22% of programs ended in an error1

1

5

5

. The Mineflayer library provided the agents with precise information about nearby blocks and creatures, even those hidden behind solid rock, along with a pathfinding plugin to plan routes4

4

. This gave the AI agents access to information unavailable to human players, making certain aspects like mining less challenging5

5

.

Finding the Stronghold and Defeating the Ender Dragon

Locating the stronghold required creative problem-solving. Agent janesick threw two Eyes of Ender from different positions and sampled each eye's flight trajectory once every 50 milliseconds, then triangulated the location through code1

1

4

4

. The calculation placed the stronghold 43 meters from the actual portal room, located 1.3 kilometers away4

4

. When auratian drowned while carrying the eyes, the team noted the location and successfully recovered them1

1

.

The final battle with the Ender Dragon consumed nearly 14 hours and resulted in 69 deaths1

1

4

4

. About 21 hours into the run, auratian discovered that ten End crystals atop pillars continuously regenerated the boss's health. The agent determined the team could reach these crystals by stacking dirt blocks1

1

4

4

. Over the next four hours and 40 minutes, all ten crystals were destroyed, and the dragon died 39 minutes later with bulicame surviving on 2.5 of 20 health points4

4

.

Comparison with Other AI Gaming Achievements

The researchers acknowledge two earlier community dragon kills but note important distinctions. One used a planner with an action-selection model on a fixed speedrun seed in Peaceful mode, while another employed a single Claude Opus 5 agent that relied on 37 ready-made, high-level tools on Normal difficulty1

1

. The Claude run completed in 3 hours and 56 minutes with 388 tool calls, only one death, and a cost of approximately $93 at API-equivalent rates1

1

.

Google DeepMind emphasizes that Gemini 4 Argon agents used only Mineflayer without game-specific tools or pre-scouted seeds, making this the first truly autonomous completion2

2

. However, the achievement comes with acknowledged limitations. The agents played on Easy difficulty, had access to information hidden from human players, and the result comes from a single run4

4

5

5

. The human speedrun record for beating Minecraft in a random world stands at under seven minutes5

5

.

Implications for Real-World Applications

Google DeepMind plans to apply this same setup for coding, science, and robotics work, according to researcher Maxence Faldor4

4

. Gemini 4 Argon is positioned as a frontier model for real-world software engineering, enterprise knowledge work including legal and finance tasks, and cybersecurity defense1

1

. The skills demonstrated—maintaining coherent plans across thousands of dependent steps for more than a day, bouncing back from mistakes, and coordinating as a team—mirror requirements for complex enterprise tasks2

2

5

5

.

Seriously, watch how these general-purpose coding agents adapt to unfamiliar environments. The fact that these AI agents are not specialized for Minecraft yet still completed the game signals progress in creating adaptable systems. Gaming benchmarks have become standard testing grounds across the AI industry, with Pokémon games capturing significant attention before Minecraft gained traction5

5

. As companies continue pushing boundaries in autonomous agent capabilities, expect more demonstrations of AI systems tackling increasingly complex, open-ended challenges that require sustained focus and adaptive problem-solving.