A team of Gemini 4 Argon agents beats Minecraft

1. Introduction

• 27 hours after its first program, the four-agent team defeated the Ender Dragon in a newly generated survival world.The run tested whether general-purpose coding agents could pursue one goal for more than a day through thousands of programs and hundreds of failures.Source: §1 · p. 2
• The agents wrote 2,783 JavaScript programs and divided the work among themselves without pre-assigned roles or a prescribed plan.They shared notes on coordinates and lessons and recovered one another’s items after deaths.Source: §1 · p. 2
• Reaching the End required dependent tasks across the Overworld, Nether, and End, including gathering resources, finding a stronghold, and destroying ten healing crystals.Prior work had trained gameplay policies or paired language models with game-specific planners, curricula, and skill libraries.Source: §1 · p. 2

2. How the agents play

• Each agent wrote and ran short programs, read the resulting text, then wrote the next program; programs ran for a median of 33.5 seconds.Players had a program running about 40% of the time, with the rest spent reading, thinking, and writing.Source: §2 · p. 2
• Mineflayer exposed player state and basic actions, but left decisions about what to do and in what order to the agents.The players shared game chat and could read one another’s notes.Source: §2 · p. 2, Figure 2 · p. 3

2.1. Agents and tools

• Five instances of the same coding agent used individual cloud machines and private and shared persistent memory; player programs ran through a command-line client.The server recorded chat, deaths, and programs, saved the world every five minutes, and restarted from its latest save after a failure.Source: §2.1 · pp. 3–4

2.2. Instructions

• Player instructions asked agents to observe and verify actions, validate tools, record lessons, and pursue the objective without exploiting evaluation flaws.The environment description explained code execution, chat, server records, and Mineflayer documentation, but contained no game strategy.Source: §2.2 · p. 4

2.3. Monitoring

• The Observer checked player claims against the record, reported findings, and could intervene only for environmental failures, not strategy.It never joined the game or ran a program in it, and flagged no specification gaming.Source: §2.3 · p. 4

3. The 27-hour run

• 27 hours and 14 minutes after the first program, the team defeated the Ender Dragon after 2,783 programs, 229 messages, and 605 deaths.The run included six server restarts and spanned about 81 in-game days.Source: §3 · p. 4

3.1. Hour 0: a team forms

• Within two hours, the team had iron tools, diamonds, and a Nether portal after quickly dividing work among the four players.Their early roles covered scouting supplies, mining diamonds, making tools, and preparing the portal.Source: §3.1 · pp. 5–6

3.2. Hours 2 to 11: into the Nether

• In the Nether, the team gathered pearls using low shelters, but bringing blaze powder back to the Overworld took more than six hours and cost items to deaths and a Ghast attack.The team turned losses into shared rules and fixes.Source: §3.2 · pp. 6–7

3.3. Hours 10 to 13: finding the stronghold

• Two Eye of Ender flight paths placed the stronghold estimate 43 m from the portal room, and bulicame recovered all twelve Eyes after auratian drowned with them.At 13 hours 17 minutes, the team reported the portal lit with all twelve eyes in place.Source: §3.3 · pp. 7, 9

3.4. Hours 13 to 27: learning the dragon

• After inspecting the dragon’s behavior and resetting it when stuck, the team discovered its healing crystals and destroyed all ten; the dragon died 39 minutes after the last fell.They reached the crystals by stacking dirt beside each pillar, punching the crystal, and digging back down by hand.Source: §3.4 · pp. 10–11

4. Emergent Behaviors

• The agents independently specialized, maintained shared memory and reusable tools, rescued teammates’ items, and corrected inconsistent reports.Their practices emerged despite instructions that did not specify roles, what to record, or how to work together.Source: §4 · pp. 12–13
• auratian retracted an announced success two minutes later, explaining that an unguarded script branch had produced the false report.gyldenia also flagged a contradiction in janesick’s report within three minutes.Source: Figure 14 · p. 13, §4 · p. 13

5. Verification

• The server log records bulicame earning the [Free the End] advancement 27 hours 14 minutes after the first program.Source: §5 · p. 14
• All 2,783 programs and their outputs, every chat message, and the full server log were recorded.Source: §5 · p. 14
• None of the programs used a server command.Source: §5 · p. 14

• About 1.35 billion input tokens and 16 million output tokens were processed over roughly 24,000 model calls.Input included 4 million thinking tokens; 91% of input tokens were served from the context cache.Source: §6 · p. 14
• About $520 was the estimated player-run cost at introductory API list prices, or $1,040 at standard prices.The estimates use list prices and session token counters, not the cost of serving the model.Source: §6 · p. 14
• The whole session cost about $810 at introductory list prices, or $1,620 at standard prices.This estimate includes the Observer agent and the day of activity after the win.Source: §6 · p. 14

• Gameplay-trained systems include MineRL (Guss et al., 2019a; Guss et al., 2019b), VPT (Baker et al., 2022), DreamerV3, STEVE-1 (Lifshitz et al., 2023), MineDojo (Fan et al., 2022), and SIMA (SIMA Team, 2024; SIMA Team, 2025).These systems learn policies from gameplay data or experience; SIMA systems act from pixels, and SIMA 2 was tested on held-out Minecraft tasks.Source: §7 · p. 14
• Language-model Minecraft work includes Voyager (Wang et al., 2024), specialized planners and skills (Wang et al., 2023; Zhu et al., 2023; Yuan et al., 2023; Li et al., 2024; Wang et al., 2025), and multi-agent studies (Gong et al., 2024; Altera.AL et al., 2024; White et al., 2025).Reported long-horizon results in this line center on Overworld technology-tree goals such as obtaining diamond tools.Source: §7 · p. 14
• Two community projects recently reported end-to-end dragon kills: one used a single agent with 37 high-level tools, and the other used a planner-action model on a fixed speedrun seed.The former was on Normal difficulty (AI Bengineering, 2026); the latter was on Peaceful difficulty.Source: §7 · p. 14
• This study gave four general-purpose coding agents standard Mineflayer and its pathfinding plugin, a one-paragraph objective, and general instructions; they wrote every program and coordinated through chat and shared notes.They played Easy survival on a random seed with no game instructions, curriculum, skill library, human-written tools, or route.Source: §7 · p. 14

• Mineflayer exposes nearby blocks and creatures, including blocks behind rock and the dragon's internal state, while its pathfinding plugin plans routes; pixel-based play remains harder.The paper notes that humans must interpret visible information or cannot see some of the exposed state, and typically navigate by eye.Source: §8 · p. 15
• Easy difficulty reduces monster damage relative to Normal or Hard, and hunger cannot kill players.Source: §8 · p. 15
• The run tests coding and coordination with a model that already knows much about Minecraft and Mineflayer, not play in an unfamiliar game.Source: §8 · p. 15
• The players died 605 times, 22% of programs ended in errors, the End fight took nearly 14 hours, and the server restarted six times.The restarts interrupted play each time.Source: §8 · p. 15
• Because this was one run—the first launch tried—the result cannot show how reliably the setup succeeds.Source: §8 · p. 15

9. Responsible development

• The agents could control only their own characters in a sandbox; they could not start processes, write files, or use server commands.Researchers could watch live and pause or stop the run, but had no cause to intervene.Source: §9 · p. 15
• An independent Observer and manual review of sampled logs found no cheating or specification gaming, and every program, message, and game event was recorded.The Observer watched for cheating or specification gaming.Source: §9 · p. 15
• The authors present visible, closely monitored coordination as one way to study agent systems before trusting them with real-world tasks.They connect Minecraft-relevant skills—team coordination, failure recovery, and pursuing a goal for more than a day—to working reliably alongside people.Source: §9 · p. 15