The typical development cycle for Agent Tuning is closed-loop: use a strong model such as GPT-4 to get the business workflow running, build training data based on prompts, fine-tune your own model, and finally replace the strong model with the trained model.