Making Agents Play NetHack
brianberns · reddit · 2026-07-18
The author built a web app that lets chatbot agents play the classic dungeon game NetHack. The results were clear: while chat models perform decently on text tasks, they clearly struggle with reading maps and understanding spatial relations.
Since a game of NetHack can last thousands of steps, the author didn't stuff the entire history into the context window. Instead, the agent maintains its own working memory, creating and deleting notes to track key states, review the past, and make short-term plans. While this method partially alleviates long-context issues, its limitations are exposed quickly—resulting in situations that are both hilarious and frustrating.
More from coding & agent
- Dev hacks reMarkable e-ink tablet to let Opus draw and animate — Stefania_druga · 2026-10-26
- Godot games are trivially decompilable — compile C# to native DLL for real protection — mflux · 2026-10-26
- Techhalla shares full 2D-to-3D workflow with Opus 5.5 and Kling 4.0 prompts — techhalla · 2026-10-26
- Jason Furman: Building Econvitals pre-AI would need ~50 people, 3 years, $20M — HarrySurden · 2026-10-26
- Dev dictates to Claude Code for 4+ hours while riding indoor bike in aero position — jacob_posel · 2026-10-26
- Proxmox and LXCs are underrated in the age of AI agents — basedjensen · 2026-10-26