Four Gemini agents beat Minecraft's Ender Dragon in 27 hours — dying 605 times
krishnan · x · 2026-10-22
A new arXiv paper shows four Gemini (Gemini 4 Argon) agents beat Minecraft's Ender Dragon in 27h 14m, dying 605 times and producing 2,783 JavaScript programs (22% of which errored).
The key insight, argues krishnan, is the failure loop: the agents kept shared notes, assigned their own roles, recovered dropped items, and converted mistakes into rules that survived context compaction — e.g., after one agent drowned holding all twelve Eyes of Ender, teammates preserved the location and despawn deadline so another could recover them.
That's the enterprise pattern for long-horizon autonomy: not a model that never fails, but failures that are bounded, observable, recoverable, and useful to the next attempt. Caveat: evidence is narrow — a single run on Easy difficulty.
More from coding & agent
- Dev sends his AI agents to fix the PHP CRUD app he wrote 8 years ago — weskambale · 2026-10-22
- Replit Agent can now incorporate your business via Stripe Atlas MCP integration — jeff_weinstein · 2026-10-22
- Using LLMs for High-Assurance Crypto: A Lean Formal Verification Experiment — jedisct1 · 2026-10-22
- Corridor's 9B fine-tune beats GPT-6 Astra and Claude Opus 5.5 on security context task — LangChain · 2026-10-22
- Claude Security adds continuous repo scanning to keep findings current — claudeai · 2026-10-22
- OpenAI upgrades all 10,000 Codex for OSS grants to Pro 200 tier — jxnlco · 2026-10-22