Abstract: To address the challenges of large-scale language models in handling complex news content and meeting multi-constraint requirements for controllable comment generation, a multi-agent collaborative approach is proposed. Initially, a comment agent endowed with subjective sentiment is constructed based on the input news text, enabling the model to interpret content and generate comments from a specific emotional perspective. Subsequently, the most suitable agent is selected from the agent repository according to the specified emotional and keyword constraints, ensuring that the output aligns with the control requirements. Finally, a review agent is introduced to filter candidate comments, thereby enhancing the coherence and relevance of the generated content to the news. Experimental results demonstrate that, compared to prompt-based learning methods, the proposed approach achieves improvements of 4.79% in ROUGE-1 and 3.51% in BERTScore. Human evaluations further confirm its significant advantages in emotional and keyword control.