To launch your first autonomous overnight run, you point your agent at the repository and let it go. Your agent reads train.py and program.md, picks a hypothesis, edits the code, and runs the training script.
Each experiment takes about minutes, so expect roughly experiments per hour and around experiments over a typical night. Karpathy himself ran about experiments over days.
While it runs, your agent writes each result to results.tsv and commits every change to git. You don't need to watch it. That's the point.