NetHack, first released in 1987, remains one of the oldest unsolved problems in artificial intelligence. Nearly four decades after its debut, no computer program has ever won the game on its current version — NetHack 3.6.6. An open project called NetHackers is now attempting to change that, and the approach is unlike anything tried before.
The challenge that outlived its creators
Winning NetHack requires descending roughly fifty procedurally generated dungeon levels, seizing the Amulet of Yendor, and escaping through five final planes. The game runs on tens of thousands of turns under permadeath, with randomized item identities and a complexity that its own developers have said takes years to master. Human experts achieve win rates around 15.9 percent, but the full player base manages just 0.4 percent.
The record of autonomous attempts is grim. In 2015, a hand-coded bot called BotHack won three games — but on an older version, 3.4.3, by exploiting a pudding-farming bug that the developers patched out that same year. Since then, on modern NetHack, every machine line has produced zero ascensions. AutoAscend, which won the NeurIPS NetHack Challenge in 2021, reached a 7.8 percent progression score and never defeated Medusa across 109,545 attempts.
Why coding agents have not cracked it yet
Over the past couple of years, coding agents that write and refine programs in iterative loops have made breakthroughs on problems that resisted every previous approach. They have solved abstract-reasoning puzzles that stalled large language models for years, beaten human winners of SAT-solver competitions, and produced new provably-correct algorithms. The pattern is consistent: a model that cannot reliably perform a task can write a program that does.
NetHack is where that pattern has not yet held. Tim Rocktäschel, co-author of the NetHack Learning Environment, marks the anniversary each year with a straightforward statement: AI still cannot learn to play NetHack. The game sits at the top of the BALROG benchmark suite as the hardest challenge, and frontier models perform worst on it by a wide margin. Whether the approach that is cracking everything else can crack this one remains the central question NetHackers exists to answer.
How NetHackers works
The project's unit of evaluation is the program itself — a deterministic bot that is cheap to run and exactly replayable. Contributors write bots using any method they choose: hand-coding, evolutionary search, or directing a coding agent to improve an existing one. Every result compounds on what came before rather than resetting with each research paper.
The project maintains a hub that tracks the best-performing bot for each of 73 starting identities and a series of milestone objectives. Anyone can pull a top-performing bot, improve it, and register the result — making one contributor's improvement available as the starting point for everyone else. The hub does not run searches or assign work. How a participant builds their bot is entirely their own.
September 2026 saw the first real movement in years. A reinforcement learning agent reached 16.98 percent progression on depth-trained policies, roughly doubling AutoAscend's previous median and pushing deeper than any prior attempt — reaching the Castle, a milestone AutoAscend never achieved. An LLM-based approach using BALROG reached 13.2 percent progression. Neither has produced an ascension, but both have cleared a bar that had been stuck for five years.
What makes the problem interesting beyond games
The research implications extend well beyond NetHack itself. A winning bot would need to convert wiki knowledge into action, decompose a tens-of-thousands-of-step objective, discover and compose reusable skills, and recover when unfamiliar seeds or stochastic events break its plan. Held-out evaluation — testing whether those skills transfer to unseen conditions — determines whether the result represents genuine generalization or a memorized trajectory.
These are not game-exclusive problems. Related loops have been applied to robotics through projects like ASPIRE, which repairs code-as-policy programs after failed rollouts, and Code as Policies, which composes perception and control into executable robot behavior. NetHack offers a faster, cheaper arena for studying long-horizon planning and skill composition without requiring a physical robot lab.
The practical entry point
Participants do not need expertise in NetHack or machine learning to contribute. A coding agent and a laptop are sufficient to start. The project runs locally, and the site describes it as genuinely addictive — a cycle of deaths and small breakthroughs where every registered improvement becomes someone else's starting point.
The project draws a deliberate contrast with the broader AI landscape. Where ARC-AGI tests how efficiently a system adapts when task evidence is deliberately scarce, NetHackers takes the opposite approach: the rules are fully public, decades of wiki knowledge are available, and every prior bot's code can be studied and built upon. Knowledge from earlier attempts is something to keep and exploit, not something to withhold. The question is not how much a system can infer from limited examples, but how much competence a process can build from all available information.
Whether that combination of open knowledge and iterative improvement can produce the first program to complete NetHack remains unknown. But after nearly four decades of silence, the number of autonomous wins on modern NetHack has moved from zero to — still zero — but the progression scores are climbing, and the community behind the effort is growing.