It will help you improve your skills and understand how to start this journey!
Ref: 1st update about this project
Tiny recap: I am using AI to write a Rust transpiler for Gameboy Advance rom to Python.
Everything changed when in Regolo.AI we got GLM 5.2 (last blogpost was 2 months ago) because before I used Minimax m2.5 and Qwen3.5-122b that doesn’t have a huge knowledge internally.
In the project there is a `docs` folder with information about GBA hardware but it isn’t enough if you need to be smart to understand a transpiled python and auto debug autonomously.
So for the same reasons I asked to GLM to check the documentation and updated it, in case it wasn’t enough and added more context with real GBA documentation again that maybe after all this time some information were lost.
With this new model in the last month everything change because also if doest’ have a huge context like the original hosted on Z.ai but only 200k it is very autonomous.
The difference it was that I had to guide Qwen3.5-122b to do things but with GLM now is enough to say “now it is time to support X.gba”, and it is faster.
It is also smarter because before I wasn’t able to use subagents to do different things so now I can split the tasks or debugging to multiple subagents with GLM as orchestrator and is uses qwen3.5-122b or itself.
An example of auto debugging, now to be faster try patching the Python generated and when see the bug fixed updates the Rust engine.
An example of the todo list that now is enforced to track what is happening.
Another example
After the previous experience I worked at every rom supported by the system an auto review of the suite and instructions to avoid getting the AI stuck always in the same issues, improving the test system but also the AGENTS.md.
This is a screen of me asking after the experience to get some other test roms working what will improve to avoid to lost times again in the same issues encountered.
Another thing I said in the previous report about this project is to work on the python code generated size, so I asked to the agents to work on this to avoid the duplication of every single assembly instructions. Initially it was doing a 1:1 transpiling but it wasn’t performant at all so now uses functions to aggregate the same stuff.
Another thing that I didn’t noticed as I was working with test roms were the assets. A ROM includes also a blob of assets that initially it was encoded in the python file itself in base64 but again it wasn’t very good.
So the next step it was to extract in a binary file all the assets from the rom and the python file to load it.
Another things I noticed that required some action it was to add some guardrails to the AGENTS.md like to not edit a .gba rom just to be a standard ROM. As test roms they need to do also dirty things and as mGBA works with them this projects has to do the same.
Those are 2 test roms for Hello World that now works!
GitHub: https://github.com/Mte90/GBAtoPy
PS: this project is running in a dedicated old laptop where I am also developing my own coding agent in Rust but this is another topic for another blogpost. Basically everyday in the office I turn on that laptop and managing by SSH with opencode while my workstation is working on other things.
Yes it is my birthday so I am happy of the actual results!
