Playing roguelike games like "slashem" is an interesting challenge for an LLMs, as this is far too complex for a single LLM prompt interaction to solve. Our approach involves chaining multiple agents and fixed tool calls via the Hatchery system.
- It's 2026. Can LLM's Play Nethack Yet?
- NetPlay agent
- Playing NetHack with Language Models - Dominik Jeurissen (YT)
Some of the initial challenges are:
- Aggressively exiting slashem during testing creates wierd conditions, some of which have no search results (e.g. "Too many hacks running now").
- Check for lockfiles in /opt/homebrew/Cellar/slashem/<version>/slashem
- Check for orphaned processes ("slashem 4")
- Watch for perm_lock problems: delete *only* perm_lock. If you delete perm, replace it with an empty file.
- The LLM has no idea how to actually play, sending inputs like "left" and "l" when trying to go left. Write a prompt for what to do.
- The LLM has no idea what things on the screen mean, or how to move to relative positions (e.g. "move to the { symbol" may go somewhere unexpected).
- The context required by screenshots is relatively large. We need to aggressively prune screen scrapes (MEMORY_DECAY=3?), or foot the bill.