Gemini 4 Argon Agent Team Clears Minecraft

Can a team of general-purpose programming agents autonomously defeat the Ender Dragon when no one assigns roles and no one writes game logic.

• Why it matters: The authors state that completing from a random spawn point requires traversing the Overworld, Nether, and End, completing a long chain of interdependent steps. The authors state that complex software engineering and enterprise tasks require the same long-horizon coding and multi-agent coordination.
• Existing approaches: The authors state that prior work either trains policies from game data or pairs language models with game-specific planners, curricula, and skill libraries. The authors state that long-horizon results reported in this line are mostly concentrated in the Overworld tech tree, such as diamond tools.
• Similar reports: The authors state that two community end-to-end dragon kills used 37 high-level tools (Normal difficulty) or a planner plus action-selection model with a fixed speedrun seed and Peaceful difficulty.
• This paper's setup: Four Gemini 4 Argon agents receive only standard Mineflayer, a pathfinding plugin, a one-paragraph goal, and general instructions; the programs are written by themselves. The authors state that, to their knowledge, this is an end-to-end autonomous completion without a pre-surveyed seed and without human game-specific tools.
• Run constraints: The world is randomly generated, survival mode, Easy, cheats disabled. Researchers can pause or stop from a dashboard; the authors state there was no reason to intervene.

Policies learned from game data

• MineRL (Guss et al., 2019a, 2019b): released human demonstrations and held a sample-efficient learning competition on the diamond task.
• VPT, DreamerV3, STEVE-1: VPT learns keyboard and mouse from online videos and crafts diamond tools (Baker et al., 2022); DreamerV3 collects diamonds from scratch with reinforcement learning (Hafner et al., 2025); STEVE-1 follows short open-text instructions (Lifshitz et al., 2023).
• MineDojo and SIMA: MineDojo calls defeating the Ender Dragon by hand in a new world one of its moonshot goals (Fan et al., 2022). SIMA and SIMA 2 act from pixels with keyboard and mouse across multiple 3D games, and SIMA 2 is also tested on held-out Minecraft tasks (SIMA Team, 2024, 2025). The authors say these systems learn policies from large amounts of game data or experience.

Language model planning and in-game multi-agent

• Voyager (Wang et al., 2024): writes Mineflayer programs into a growing skill library under an automatic curriculum. This line also builds on interleaved reasoning and tool use (Yao et al., 2023) and code as policy (Liang et al., 2023).
• Specialized modules: DEPS, GITM, JARVIS-1, Plan4MC, Optimus-1 add Minecraft-specific planning, knowledge, memory, or learned skills (Wang et al., 2023; Zhu et al., 2023; Wang et al., 2025; Yuan et al., 2023; Li et al., 2024). The authors say the long-horizon results reported in this line focus on the overworld tech tree.
• Coordination research: MindAgent, Project Sid, Collaborating Action by Action examine coordination and social behavior in games, including Minecraft (Gong et al., 2024; Altera.AL et al., 2024; White et al., 2025). The paper only describes these three.

Community end-to-end dragon kills

• Mine AI MCP (AI Bengineering, 2026): a single Claude Opus 5 calls 37 high-level Minecraft tools, on Normal difficulty.
• Astra and JEV (Malde, 2026): a planner with an action-selection model, fixed speedrun seed, Peaceful difficulty.
• Authors' distinction: The difference is mainly in what the agents receive. This paper is four general-purpose coding agents, stock Mineflayer and a pathfinding plugin, a one-sentence goal, and general instructions. The authors say no additional game instructions, curricula, skill libraries, human-written tools, or routes were added; random seed, Easy, coordination through chat and shared notes, and all programs written by themselves.

Baselines: The paper has no same-run baselines. The task environment is a randomly generated Minecraft Java Edition 1.20.4 survival world. The diamond tasks, held-out tasks, and the dragon-kill setups in related work are used to illustrate differences, not as controlled experiments.

The authors place this paper alongside the above work: the test is whether general-purpose coding agents can turn existing knowledge into programs and team coordination over a day, rather than learning another game policy or being fitted with another set of human game tools.

Four Gemini 4 Argon agents, without viewing the screen, repeatedly write short JavaScript programs executed by Mineflayer in the same survival world, collaborating via chat and shared memory. A fifth Observer only witnesses and does not participate.

World and Goal

• Environment: Minecraft Java Edition 1.20.4, survival mode, Easy, randomly generated world, cheats disabled, no human input during the run.
• Goal: The game-related instruction is one paragraph: defeat the Ender Dragon from the spawn point of the shared survival world in that session; this is the only criterion; any means within survival rules are allowed, and the players decide how to divide the work themselves. There is also a sentence stating they are a four-person team.
• General instructions: Observe before acting and verify results; validate new tools on short tasks first; build your own tools when manual work would limit speed or reliability; record mechanisms, strategies, and workarounds in shared memory; continue working while background tasks run; be honest about actions and results; pursue the goal rather than exploiting the environment or evaluation flaws.
• Environment description: How to run code with the bot, how chat works, where the server records the game, and where the Mineflayer documentation is. The authors state it contains no game strategy.

Action Interface

• Perception and action: Programs can read position, health, inventory, and surroundings; can walk to a point, mine or place blocks, craft, and attack. Mineflayer does not include gameplay; the sequence is decided by the agents. They do not view pixels and do not use a keyboard or mouse.
• Loop: Write a program, run it until the task is complete or the agent's self-chosen time limit, then read the returned text. The client default wait cap is 60 seconds. The median program is 33.5 seconds. Each player has a program running about 40% of the time; the rest is reading results, thinking, and writing the next program.
• Sandbox: Programs are async function bodies with the pathfinding plugin loaded and no extra helpers. Node.js built-in modules and installed Mineflayer packages can be loaded; processes cannot be started and files cannot be written. Values placed on the global object persist across programs.
• Server: Records chat, deaths, and progress; saves every program and its output; saves the world every five minutes; restarts from the most recent save after a failure.

Agents, Memory, and Network

• Instances: The four players and the Observer are the same general-purpose programming agent, using Gemini 4 Argon, each on its own cloud machine. Tools include a versioned code workspace, shell, file editing, web search, and persistent memory with a private section plus a team-shared section.
• Compression: Context is compressed when too long, the same as doing software work. The instructions require writing content that must survive compression into memory, such as ongoing work and next steps.
• Network usage: None of the 2,839 programs run by the players during the session (2,783 before winning) made network requests or cited web sources. Chat and Observer reports also did not mention such requests.

Collaboration Channels

• Chat: Messages sent via game chat are received by all bots and appear in the output of that player's next program.
• Notes: Each agent keeps notes that others can read. Figure 2 marks chat and notes as shared among the four players.
• Unspecified parts: The instructions do not mention roles, chests, or rescue, except that the players decide how to divide the work themselves. The general instructions require reusable tools, shared notes, and honest reporting, but do not specify what to build, record, or how to work together.

Witnessing and Adjudication

• Observer: Independently witnesses and never plays the game. The instructions require checking against records before reporting player claims, reporting specification gaming and fabricated results, and not altering the game; it intervenes only on environment failures, such as broken tools or quota limits, and never on strategy.
• Reports: It publishes real-time reports with progress summaries, verified events, and infrastructure findings. The authors state it did not flag specification gaming and never joined the game or ran programs in it.
• Adjudication: Completion is recorded in the server log, i.e., obtaining the advancement [Free the End]. Section 5 states that no recorded program used server commands. Researchers can pause or stop; the authors state there was no reason to intervene.

Experimental setup

• Subjects: Players janesick, gyldenia, auratian, bulicame, each driven by Gemini 4 Argon. An Observer of the same model does not play.
• Environment: Minecraft Java Edition 1.20.4, Survival, Easy, random seed, cheats off, no human input during the run. The paper does not give the seed value.
• Comparison: The paper does not report same-run baselines or repeated runs. The authors say this is the startup of the first attempt.
• Duration composition: From first program to completion spans about 81 game days; 6 server restarts, totaling about 20 minutes of pause.
• Metrics: Time to the log progress [Free the End]; number of programs, unique JavaScript lines, chat messages, deaths and causes, and proportion of programs ending in errors.
• Verification: The Observer verifies claims against records. Section 9 says the authors manually reviewed a sample of logs. The paper does not report the sample proportion, number of repetitions, or agreement with the Observer.

Main results

Result	Value	Source
Completion time	27 h 14 min	Sections 3, 5, Figure 1
Programs before victory	2,783	Section 3
Unique JavaScript lines	63,313	Intro statistics item
Chat messages before victory	229	Section 3
Deaths before victory	605	Section 3, Figure 13
Programs ending in errors	22%	Section 8

• Authors' summary: The authors say victory required maintaining a coherent plan over more than a day and thousands of dependent steps. Section 3 gives human references: HowLongToBeat main story median 45 hours, random-seed speedrun record under 7 minutes; and states the agents are four, on Easy, acting through programs rather than screen and keyboard.
• The End: The authors say the team slowed down in the End. The dragon caused 69 deaths, and the End fight lasted nearly 14 hours. Before the last crystal was destroyed, program readings never went below 198.5; the dragon died 39 minutes after the crystal fell, with bulicame at 2.5/20 health.
• Role labels: Figure 1 text labels tools, stronghold, and stronghold on janesick; supplies and six crystals on gyldenia; the Nether portal and crystal plan on auratian; the Enderman shelter and final blow on bulicame.

Milestones

• Early phase: Figure 4 marks stone tools +0 h 06, iron tools and bucket +0 h 14, flint and steel +1 h 04, Nether portal +1 h 42, Nether fortress +2 h 57, Ender pearl +3 h 06, blaze rod +4 h 17. The text says iron tools, diamonds, and a portal existed within two hours; janesick placed the last obsidian block, and auratian lit the frame two seconds later.
• Locating and the End: Eye of Ender +10 h 43, stronghold +13 h 17, entering the End +13 h 33, ten crystals +26 h 34, dragon +27 h 14 (Figure 4). janesick threw twice, sampling every 50 milliseconds, with the intersection 1.3 km from the village and 43 m from the actual portal room (Figure 7). The text says all ten crystals were destroyed over the following 4 hours 40 minutes.
• Rescue: At 5.5 hours, the shared chest had 15 pearls. When auratian drowned carrying all 12 eyes, 8 m from bulicame; bulicame picked them all up two and a half minutes before despawn. At 13 h 17 min, bulicame reported the End portal was lit.

Emergent behaviors

• Messages: Figure 13 pattern-matches the 229 messages before victory, with one message possibly matching multiple: exact coordinates 173, shared chest 64, warnings 24, item rescue 20, experience and tips 11.
• Tools: The authors say roles formed themselves. Within 19 minutes of the start, gyldenia warned about a Mineflayer crafting quirk; a helper with repair appeared in 655 programs written by the four agents. 146 pieces of state and code were retained across programs, including 42 helpers; a counter that made leftover timed-out code stop itself was referenced by 1,519 programs.
• Corrections: auratian's program announced a success that had not happened, and corrected itself two minutes later. janesick reported bringing back 12 blaze powder one second after burning to death, and gyldenia pointed out the contradiction within three minutes.

Verification and costs

• Verification: Section 5 says the victory is in the server logs, the progress was earned by bulicame, at 27 h 14 min after the first program. None of the recorded programs used server commands. Section 2.3 says, per the Observer's and authors' manual review of the logs, there were no safety incidents.
• Tokens: As of victory, the four players used about 1.35 billion input tokens, 91% served by context cache; about 16 million output tokens, of which 4 million were thinking tokens; about 24,000 model calls (Section 6).
• List prices: At public API list prices, entry price about $520, standard price $1,040. Including the Observer and the day after victory, the full session is about $810 ($1,620). The authors say this is an estimate from list prices and token counts, not the cost of serving the model.

Other ablations and analyses

The paper does not report ablations or repeated experiments. Other counts are as follows:

• Death causes, Figure 13: mobs 351, dragon 69, drowning 67, fall and void 57, lava and fire 47, other 14.
• Programs and messages, Figure 1: auratian 735 / 53, gyldenia 727 / 79, bulicame 675 / 47, janesick 646 / 50.
• Crystals, Figure 12: gyldenia 6, auratian 3, bulicame 1, janesick 0.
• Healing: the text says at about 21 hours auratian noted that 10 crystals kept the dragon at full health; Figure 11 chat text reads 200/200 at that time, with perch damage capped at 50 HP before forced takeoff.
• Reset: at 17.5 hours, the four players deliberately jumped into the void within six minutes, so that the empty End would unload and reset the stuck dragon.
• Failures during the process: Section 8 writes 605 deaths, 22% of programs ending in errors, the End fight lasting nearly 14 hours; Section 3 writes 6 server restarts.

Limitations and future directions noted by the authors

• Code perception and pathfinding: Mineflayer reports surrounding blocks and entities, including blocks behind solid rock and the dragon's internal state; a pathfinding plugin plans routes. The authors say playing from pixels with keyboard and mouse, as in SIMA 2, remains much harder (Section 8).
• Easy: Mobs spawn and attack, but deal less damage than on Normal or Hard, and hunger cannot kill the player (Section 8).
• Prior knowledge: Gemini 4 Argon knows a great deal about Minecraft and Mineflayer from training data. The authors say this test is whether that knowledge can be turned into runnable programs and day-long team coordination, not whether it can play a game never seen before (Section 8).
• Failures along the way: 605 deaths, mostly from mobs; 22% of programs ended in errors; the End fight lasted nearly 14 hours (Section 8).
• Infrastructure: The world server restarted 6 times, each interrupting the game (Section 8).
• Single run: The authors say they describe the startup of the first attempt, which cannot show how stable this setup's success is (Section 8). The paper does not list a separate Future Work section.

Experimental coverage

• The subjects are four players of the same Gemini 4 Argon and one non-playing Observer; the environment is Minecraft Java Edition 1.20.4 Survival, Easy, random seed, cheats off (Sections 2, 3).
• Actions are writing JavaScript with text returned by Mineflayer; median program 33.5 seconds, default wait cap 60 seconds, with time limits chosen by the agent; each player has a program running about 40% of the time (Section 2).
• Records include all programs and outputs, chat, and server logs; completion is defined by [Free the End]. The Observer and authors reviewed for specification gaming and found none reported. The paper does not report the sample size of human review or agreement with the Observer (Sections 2.3, 5, 9).
• Costs are estimated from public API list prices and session token counts; the authors say this is not the cost of serving the model (Section 6).
• Section 7 compares MineRL, VPT, DreamerV3, STEVE-1, MineDojo, SIMA/SIMA 2, Voyager, DEPS, GITM, JARVIS-1, Plan4MC, Optimus-1, and two community dragon-kill reports. The paper does not report head-to-head scores for these systems.

Four general-purpose coding agents in a sandbox Minecraft survival world, with defeating the Ender Dragon as the only goal, write their own programs and coordinate with each other.

• Question: The authors ask whether coding agents not built for Minecraft can pursue the same goal over more than a day, thousands of programs, and hundreds of failures, and coordinate with teammates.
• Method: Gemini 4 Argon writes JavaScript via Mineflayer and reads text returns, collaborating through in-game chat and shared memory. There are no preset roles, curricula, skill libraries, or human game tools. An Observer verifies against records, does not participate, and does not flag specification gaming.
• Result: Easy, random seed, single run. 27 h 14 min after the first program, 2,783 programs, 229 messages, 605 deaths, server logs record bulicame earning [Free the End]. 63,313 unique JavaScript lines. 22% of programs ended in errors; the dragon caused 69 deaths, and the End fight lasted nearly 14 hours.
• Collaboration: Roles were self-proposed about 10 minutes in; losses were written into shared rules; 12 Eyes of Ender were recovered by coordinates and despawn timers. The intersection of two eye throws was 43 m from the portal room. Before victory, the four players used about 1.35 billion input tokens, at entry list price about $520.
• Authors' conclusion: The authors say that to their knowledge, using only standard Mineflayer, without pre-surveying the seed, and without human-written game-specific tools, this is an end-to-end autonomous completion. The authors also state Easy difficulty, code perception, game knowledge from training data, numerous failures, and that this is the startup of the first attempt, which cannot show how stable this setup's success is. The authors say team coordination, recovery from failure, and pursuing a goal over a day are also capabilities needed when agents work alongside humans.