DatawiseAgent is a notebook-centric LLM agent framework for data science automation. To overcome the narrow task scope, limited generalization, and dependency on state-of-the-art LLMs of existing LLM agents, we draw inspiration from the work of human data scientists and introduce a unified interaction representation and an FST-based multi-stage architecture. DatawiseAgent enables flexible long-term planning, incremental solution development, and recovery from execution failures, achieving state-of-the-art performance that surpasses robust foundational models like AutoGen and TaskWeaver across a variety of data science scenarios. Furthermore, it demonstrates robustness and scalability, demonstrating stable performance degradation even with weaker models.
