Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search | ScienceToStartup