Prompts within a complex system are brittle

I have seen some teams be successful by replacing prompts with smaller, more deterministic components and improved reliability with fine-tuning. Anyone else have success with this approach?

Seems to help a lot with agents