What's this about?
GPT-6 Astra played NetHack through a terminal, wrote its own guarded interface and persistent notes, repaired its tooling mid-run and completed the game after 37,140 turns on its third try.
What exactly is a harness?
It is the software surrounding a model: tools, state parsers, memory and safety checks. Here it converted terminal screens into manageable observations and stopped dangerous batches when health, position or nearby threats changed.
What's the community saying?
Ethan Mollick's discussion post drew 225 likes; the author published all three runs and server records.
- The striking part is long-horizon execution and self-built tooling, not merely knowing NetHack rules.
- Skeptics note unrestricted spoilers, source access, evolving instructions and roughly $300 in extra credits plus a subscription.
- One success in three attempts establishes possibility, not a reproducible win rate or an apples-to-apples benchmark.