Data augmentation is widely adopted in molecular graph learning due to its practical benefits in enhancing model performance. However, conventional augmentation like edge deletion often destroys critical chemical semantics, as molecular properties are highly sensitive to structural changes. To address this, we introduce CMAS, a novel  ̲Collaborative  ̲Multi- ̲Agent  ̲System for enhancing data augmentation in molecular graph classification. Specifically, for each input molecule, we first use a BRICS-based decomposition strategy with prompt-guided LLM generation to produce structurally diverse candidates while preserving the core scaffold. To ensure that the generated candidates preserve the physicochemical properties of the original molecule, we retrieve domain-specific chemical knowledge from a knowledge graph and prompt an LLM to assess property consistency, effectively filtering out semantically divergent samples. Finally, we perform rule-based sanitization with iterative feedback to ensure chemical validity and preserve distributional diversity, resulting in a high-quality augmented molecular set. Extensive experiments on MoleculeNet benchmarks validate the superiority of CMAS against several state-of-the-art baselines. Code is available at https://github.com/Catfromuibe/CMAS.