“Running Continual Harness: Self-Improving AI Agent for Pokémon Emerald (Full Setup Guide)”

I am trying to run Continual Harness — the reset-free, self-improving Pokémon agent from the May 2026 research paper — on Pokémon Emerald using Gemini 3 Pro.

What I’ve Accomplished So Far:

  • Successfully cloned the official GitHub repo into ~/continual-harness
  • Fixed several package issues:
    • Installed and verified OpenCV (cv2)
    • Upgraded from the old google.generativeai to the new google-genai SDK
    • Installed FastAPI, Uvicorn, and other web server dependencies
  • Created a dedicated conda environment called continual-harness
  • Made a backup of PokeAgent.py and tried fixing import errors
  • Got the agent to partially initialize (it loads system prompts, tools, and the HarnessEvolver)

Current Status:

  • I’m using Python 3.12 in the conda environment
  • The main blocker right now is mgba==0.10.2 — the GBA emulator has no compatible wheel for Python 3.12, so requirements.txt fails to install completely.

Goal:

Get the full Continual Harness agent running on Pokémon Emerald so it can play the game while continuously improving its own harness (prompts, skills, strategies, etc.) through o-learning.