This paper introduces QuantAgent, a novel framework for developing autonomous trading agents based on Large Language Models (LLMs). The core innovation is a principled two-layer loop architecture that addresses the fundamental challenge of building domain-specific knowledge bases for quantitative investment applications. In the inner loop, the agent refines its responses by drawing from its existing knowledge base, while in the outer loop, these responses are tested in real-world scenarios to automatically enhance the knowledge base with new insights.

The authors demonstrate that this approach enables the agent to progressively approximate optimal behavior with provable efficiency, representing a significant advancement in specialized LLM agent development. Through the instantiation of this framework as QuantAgent, the paper shows practical success in mining viable trading signals and improving financial forecast accuracy. The research bridges the gap between general-purpose LLM capabilities and specialized quantitative finance applications, offering a systematic approach to autonomous agent learning in complex financial domains.