This paper introduces a groundbreaking three-stage framework that leverages Large Language Models (LLMs) within a risk-aware multi-agent system for automated strategy discovery in quantitative finance. The research addresses the fundamental brittleness of traditional deep learning models in financial applications by developing a comprehensive approach that combines prompt-engineered LLMs for generating executable alpha factor candidates from diverse financial data sources, implementing a sophisticated multimodal agent-based evaluation system that filters factors based on market status and predictive quality while maintaining category balance, and deploying dynamic weight optimization mechanisms that adapt to changing market conditions.

The methodology represents a significant advancement in quantitative finance by extending LLM capabilities beyond natural language processing to quantitative trading applications. The framework's experimental validation demonstrates exceptional performance, with the system achieving a remarkable 53.17% cumulative return on the SSE50 index from January 2023 to January 2024, significantly outperforming all established benchmarks. The research shows robust performance across both Chinese and US market regimes, with particular strength in risk-adjusted performance and downside protection. This work provides a scalable architecture for financial signal extraction and portfolio construction that bridges the gap between advanced language models and practical quantitative investment applications.