Definition
Plain language
A famously punishing text-based dungeon-crawler game used to test whether AI agents can survive very long, complex play-throughs.
As stated in the literature
A roguelike game used as a long-horizon RL benchmark; episodes can run to ~100,000 steps with deep item and dungeon complexity, making it a stress test for agent memory and planning.
Why it matters: Its extreme length and complexity push agents far beyond short tasks, exposing how well they can plan and remember over very long stretches.
For example, an agent in NetHack might have to survive tens of thousands of turns, managing food, weapons, and monsters deep in a shifting dungeon.
Heard on the show
“In NetHack, the roguelike that runs to a hundred thousand steps, the starting agent keeps a file called dungeon map, and it's append-only.”Episode 192 — A 32B Open Model Matched Frontier Systems By Learning to Take Notes