You don't write training code anymore. Your agent does that. Your job is to write program.md, the plain Markdown file your agent reads at the start of every experiment cycle.
This is a shift. Your output is no longer Python. It's English. You describe what to optimize, what constraints to respect, and what strategies to consider. Your agent translates that into code changes, runs them, and measures the results. The quality of your program.md directly determines the quality of your overnight run.