Four Gemini 4 Argon agents beat Minecraft in a 27-hour autonomous coding run
A paper reports four Gemini 4 Argon agents wrote 2,783 programs and died 605 times over 27 hours before defeating Minecraft's Ender Dragon.
A paper posted to arXiv on October 15 reports that four Gemini 4 Argon agents finished Minecraft on their own. Each agent controlled a player through the open-source Mineflayer library and coordinated through in-game chat and shared memory. Nobody assigned roles or wrote game logic. Starting from a random world on Easy difficulty, the team wrote 2,783 programs and died 605 times. It defeated the Ender Dragon after 27 hours and 14 minutes. The authors say this is, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents. They frame it as a test of the long-horizon coding and multi-agent coordination that real software and enterprise tasks need.
For anyone building long-running agents, the useful details are in the process. Crypto Briefing, working from the paper, reports the agents split up responsibilities within ten minutes. It says they kept shared notes on locations and lessons from failures, built 42 reusable helper functions, and survived six server restarts. It also reports about 24,000 model calls and 1.35 billion input tokens, and estimates the run cost about $520 at introductory API pricing. That cost figure is the outlet's estimate, not one we confirmed in the paper.
The setup also has clear limits. Digg's summary notes the agents acted through code, not pixels, and a pathfinding plugin planned their routes. It reports that 22% of programs ended in errors. A fifth Observer agent and a manual log review found no cheating, but that is the team's own account, not an external replication. It's also unclear whether the code or run logs are public.