Google DeepMind's Gemini 4 Argon agents defeat Minecraft's Ender Dragon in 27-hour autonomous run

Four AI agents wrote nearly 2,800 programs themselves, marking a milestone in long-horizon autonomous task completion

By LineZotpaper

Published20 October 2026

Read Time3 min

Four AI agents powered by Google DeepMind's Gemini 4 Argon model have autonomously defeated the final boss of Minecraft in 27 hours and 14 minutes, writing 2,783 programs of their own in the process. The feat, described in a research paper as the first end-to-end autonomous completion of the game by AI agents using only open-source bot infrastructure, cost an estimated $520 in tokens at introductory pricing.

Google DeepMind has reported that a team of four Gemini 4 Argon agents successfully completed Minecraft by killing the Ender Dragon, the game's final boss, in a 27-hour and 14-minute autonomous run. The agents, operating in a shared survival world on Easy difficulty, wrote all of their own JavaScript programs to coordinate tasks such as gathering resources, exploring the Nether, locating the End portal, and executing the final battle.

The agents, each controlling a player through the open-source JavaScript library Mineflayer, communicated via in-game chat and shared notes. According to the research paper, there was no human input during the run, and no roles were assigned or game logic written by people. Within about 10 minutes of starting, the agents spontaneously proposed a division of labor in chat: one took on food, beds, and flint; another focused on diamonds; a third on iron and tools; and the fourth on building the Nether portal.

During the run, the agents died 605 times, including 69 deaths during the dragon fight. About half the total time was spent fighting the Ender Dragon. The agents initially struggled with the boss's ten crystals, which kept regenerating its health, until one agent discovered they could use dirt pillars to reach and destroy the crystals. After four hours and 40 minutes of crystal destruction, the dragon was killed 39 minutes later.

The run used an estimated 1.35 billion input tokens, with 91 percent served from cache, and about 24,000 model calls generating 16 million output tokens. Google's introductory pricing put the cost at roughly $520; at normal pricing, the authors estimated it would exceed $1,000.

The authors note that 22% of the agents' programs ended in an error, though the run ultimately succeeded. They highlighted that the demonstration was designed to showcase the model's capability for complex, long-horizon workflows and agent coordination, rather than game-playing skill per se. Google's blog describes Gemini 4 Argon as a frontier model for real-world software engineering, enterprise knowledge work, and cybersecurity defense.

The paper acknowledges earlier community efforts to defeat Minecraft's dragon using AI, including a planner on a fixed speedrun seed in Peaceful mode and a single Claude Opus 5 agent that used 37 pre-made, high-level tools on Normal difficulty. That Claude run took 3 hours and 56 minutes with 388 tool calls and one death, costing about $93 at API-equivalent rates.

Analysis

Why This Matters

This is a significant demonstration of AI agents autonomously planning, negotiating, and executing a complex, multi-hour task without human guidance or specialized game logic.
The approach—using general-purpose coding agents rather than game-specific models—suggests potential applications in software engineering, enterprise workflows, and other domains requiring sustained coordination.
The cost comparison (hundreds to thousands of dollars per run) highlights the current expense of frontier-model autonomy, raising questions about economic viability for real-world deployments.

Background

Minecraft has long been a benchmark for AI research because of its open-ended, survival-oriented gameplay requiring exploration, resource gathering, and strategic planning. Previous AI achievements in the game have ranged from simple task completion to defeating the dragon on specific seeds with pre-built tools. Google DeepMind's approach differs by having the agents write their own programs from scratch and coordinate via chat, mirroring human gameplay more closely. The research builds on a broader trend of AI agents playing games like Pokémon Red and World of Warcraft, using language models operating on text-based observations rather than screen captures.

Key Perspectives

Google DeepMind researchers: The run demonstrates that frontier models can handle complex, long-horizon workflows and agent coordination for real-world tasks, not just game-playing. The agents' ability to self-organize and recover from failures (such as drowning and losing items) is a key capability.

AI efficiency and cost skeptics: At over $500 for a single run (and potentially double that without promotional pricing), the approach remains economically impractical for most applications. The 22% program error rate and 605 deaths also suggest significant room for improvement in reliability.

Competing AI labs and open-source community: Earlier, cheaper approaches (such as the Claude Opus 5 run at $93 with only one death) may argue that specialized, smaller models with better tool integration can outperform general-purpose agents at lower cost. The community may also question the novelty, given prior human-assisted AI completions.

What to Watch

Whether Google DeepMind publishes further detail on how the agents' self-written programs were debugged and reused across the run, and whether the approach generalizes to other games or real-world coordination tasks.
The cost trajectory: as model inference gets cheaper, the economic case for autonomous multi-agent systems may shift dramatically.
The response from other AI labs (such as Anthropic and OpenAI) who have published competing game-playing runs, potentially sparking a public benchmark for agent autonomy.

Sources

Google DeepMind claims four AI agents beat Minecraft in 27 hours by writing 2,783 programs of their own · Latest from Tom's Hardware

Checked

12 claims in this story checked against 1 source: 10 supported.

Show each claim

SupportedThe agents defeated Minecraft’s Ender Dragon in an autonomous run lasting 27 hours and 14 minutes.Latest from Tom's Hardware: “The run, starting with the first program and ending with the dragon's death, took 27 hours and 14 minutes.”
Partly supportedThe research paper describes the run as the first end-to-end autonomous AI completion of Minecraft using only open-source bot infrastructure.Latest from Tom's Hardware: “The authors say this is the “first end-to-end autonomous completion of Minecraft by AI agents,” at least “to our knowledge,” qualified by the use of Mineflayer “without pre-scouted seeds or other game-specific tools written by humans.””
SupportedThe agents wrote 2,783 programs themselves.Latest from Tom's Hardware: “The game of Minecraft is no match for a team of four Gemini 4 Argon agents, as they were able to defeat it in 27 hours with 2,783 self-written programs”
Not in the sourcesThe run marked a milestone in long-horizon autonomous task completion.Latest from Tom's Hardware
SupportedThe agents’ run cost an estimated $520 at introductory pricing.Latest from Tom's Hardware: “Google’s Minecraft run, using Gemini 4 Argon’s 50% introductory pricing, which runs from initial API access that followed the model announcement through Nov. 30, cost more than half a grand”
SupportedThe agents operated in a shared survival world on Easy and coordinated through in-game chat and shared notes.Latest from Tom's Hardware: “The playthrough involved a random, shared survival world on Easy with four agents”
SupportedThere was no human input during the run.Latest from Tom's Hardware: “There was “no human input during the run””
SupportedThe agents died 605 times, including 69 deaths during the dragon fight.Latest from Tom's Hardware: “The agents did die 605 times”
SupportedAbout half the run was spent fighting the Ender Dragon.Latest from Tom's Hardware: “about half the run was spent fighting the Ender Dragon”
SupportedThe run used about 1.35 billion input tokens, 91% served from cache, and around 24,000 model calls involving about 16 million output tokens.Latest from Tom's Hardware: “Google’s run took about 1.35 billion input tokens with 91% served from cache. About 24,000 model calls involved around 16 million output tokens”
SupportedThe agents’ programs ended in an error 22% of the time.Latest from Tom's Hardware: “With Minecraft, 22% of the agents’ programs ended in an error”
SupportedThe earlier Claude Opus 5 run used 37 pre-made high-level tools, took 3 hours and 56 minutes, and had one death.Latest from Tom's Hardware: “a single Claude Opus 5 agent that used 37 ready-made, high-level tools on Normal”