The paper introduces the *Iterative Experience Refinement* (IER) framework, which enables large language model (LLM)-powered agents to iteratively refine their experiences during task execution. The authors propose two fundamental patterns for experience refinement: the successive pattern, which refines based on the nearest experiences within a task batch, and the cumulative pattern, which acquires experiences across all previous task batches. To manage the experience space effectively, a heuristic experience elimination mechanism is introduced, prioritizing high-quality and frequently used experiences. Extensive experiments demonstrate that while the successive pattern may yield superior results, the cumulative pattern provides more stable performance. Additionally, experience elimination allows achieving better performance with just 11.54% of a high-quality subset of experiences. The IER framework enhances the adaptability and efficiency of LLM agents in software development tasks, fostering a paradigm shift in the design of LLM agents.The paper introduces the *Iterative Experience Refinement* (IER) framework, which enables large language model (LLM)-powered agents to iteratively refine their experiences during task execution. The authors propose two fundamental patterns for experience refinement: the successive pattern, which refines based on the nearest experiences within a task batch, and the cumulative pattern, which acquires experiences across all previous task batches. To manage the experience space effectively, a heuristic experience elimination mechanism is introduced, prioritizing high-quality and frequently used experiences. Extensive experiments demonstrate that while the successive pattern may yield superior results, the cumulative pattern provides more stable performance. Additionally, experience elimination allows achieving better performance with just 11.54% of a high-quality subset of experiences. The IER framework enhances the adaptability and efficiency of LLM agents in software development tasks, fostering a paradigm shift in the design of LLM agents.