When you iterate on program.md between runs, include specific results from previous experiments. Your agent reads this before every cycle.
Example:
"From previous runs:
- Halving batch size: val_bpb to (keep)
- Adding th layer: val_bpb to (keep)
- th layer: val_bpb regressed (discard)
- Current best: val_bpb "
This saves your agent from re-testing ideas that already failed. Without this context, your agent might spend experiments rediscovering that a th layer doesn't work. With it, your agent skips straight to untested ideas.