Tuberculosis (TB) remains a global health crisis, necessitating accurate prediction models to guide resource allocation and interventions. Traditional time-series and machine learning models often lack flexibility and precision in capturing complex TB trends. This study introduces TB-DeepSeek-Agent, a novel framework integrating a large language model (DeepSeek) with an Agent-based architecture to enhance TB incidence forecasting. The model leverages historical data, memory mechanisms, and reflective thinking to refine predictions dynamically. Evaluated on 135,802 TB cases from five Chinese cities (2011–2021), TB-DeepSeek-Agent achieved superior performance, with a 100% acceptance ratio (5% error margin), the lowest deviation ratio (0.0129), and RMSE (30.5876), outperforming traditional models (e.g., ARIMA, LSTM) and standalone LLM-based approaches. A web-based application was developed to automate simulations and generate actionable reports, streamlining TB research and policymaking. This work demonstrates the potential of Agent-enhanced AI models in public health analytics, offering a cost-effective solution for TB trend prediction and resource optimization.