How to quickly identify high-quality frontier topics from massive scientific research data to assist researchers in accurately carrying out scientific research work is of great importance. Traditional analysis methods have some bottlenecks, such as weak cross-domain adaptability, high resource consumption and low efficiency. In order to solve the above problems, a frontier topics mining method via AI-agent is proposed. A generative-verification dual-agents (D-Agents) architecture is innovatively constructed. Firstly, prompt engineering is used to construct generative agent (G-Agent), and the semantic understanding ability of large-scale pre-trained language models is used to realize the automatic generation of candidate frontier topics; Then, the verification agent (V-Agent) is introduced to establish a multi-dimensional evaluation system, and the candidate results are automatically verified from the dimensions of academic novelty, topic accuracy and completeness to identify frontier topics. The effectiveness of the proposed method is verified by constructing three labeled test dataset including computer vision (CV), natural language processing (NLP), and machine learning (ML). The experimental results show that D-Agents can be competent for frontier topics mining tasks in multiple domain at the same time. On three manually labeled datasets: CV-DataSet, NLP-DataSet and ML-DataSet, the accuracy rate of D-Agents exceeds 74% while maintaining the coverage rate of more than 85%. Compared with traditional bibliometric methods, the accuracy and coverage rate of frontier topics mining in three different fields: altitude sickness, recommendation system and oyster reef ecosystem have reached more than 67%. It can effectively alleviate the hallucination problem of G-Agent through the automatic generation and self-verification mechanism in D-Agents, and greatly improve the efficiency of frontier topics mining.