Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use