Description

An agent (backed via an API key) that can play NetHack in a tmux/screen-like session that you can watch.

Admittedly not looking for collaboration right away, since I feel I first need to learn more about how to do this before I could effectively collaborate with others, and this is about exploration. Maybe later in the week or post-HW?

Goals

• Understand how to program and develop an agent
• How to have the agent interact in a chaotic world with lots of freedom
• How to provide initial prompt, augmented by persistent and session memory
• How to enable an agent to parse "natural" context from the play's screen
• How to augment this with resources a human would have access to (the NetHack Wiki, in game help)

• Mainly have fun revisiting my childhood's most played game in a modern context

• I'm somewhat hopeful that at the end of the week, I should have an agent playing NetHack badly.

Stretch goal:

• New LLM benchmark? :-)

Resources

• Probably going to use OpenRouter as an abstraction so I can easily compare models
• Codex/Gemini CLI/Claude Code for development of the agent itself
• Undecided between python or Go