An LLM Beat NetHack
2 days ago
- GPT-6 Astra achieved the first recorded NetHack ascension by an LLM agent, playing as a dwarven Valkyrie named CodexDelver.
- The agent built its own harness, tools, and memory system without human coding involvement.
- Architecture included terminal observations, checked batching for movements, and task-specific helpers for route planning and inventory.
- External memory and access to spoilers, wiki, and source code were used, but the game state was only from terminal display.
- The winning run took from September 9 to 21, 37,140 turns, with harness modifications made by the agent during gameplay.
- Cost was approximately $200/month subscription plus $300 in additional credits.
- The result is verifiable through server dumps, local audit records, and GitHub archives.