Teaching Gemma to Code

A handbook for turning Gemma 4 into a coding agent, measured run by run

Why this handbook

My advisor suggested I enter the Gemma 4 Developer Agent competition. The task: build an agent on Gemma 4 that fixes real GitHub issues, scored by whether the fix passes that issue’s tests.

I use coding agents every day, but I wanted to know how one works inside, and to measure it instead of guessing. So I’m writing down what I learn, chapter by chapter, with the code and the scores next to it.

Every number in this handbook comes from a run I can point to in the experiment log. When something didn’t work, I say so.