Agentic data science patterns

Agentic data science patterns transition static notebooks into autonomous loops where LLMs iterate on feature engineering, model selection, and error analysis.

Modern data science is shifting from manual experimentation to agentic workflows using frameworks like LangGraph and AutoGen. These patterns employ specialized agents to handle discrete tasks: one for SQL generation, another for cleaning outliers, and a third for hyperparameter tuning. By implementing iterative loops (think ReAct or Plan-and-Execute), these systems can self-correct when a Python script fails or a model underperforms. This approach reduces the time to find optimal XGBoost parameters or valid feature sets from hours of human labor to minutes of automated execution. The focus is on robust orchestration: ensuring the agent can validate its own statistical assumptions before finalizing a deployment pipeline.

https://www.deeplearning.ai/the-batch/how-agents-can-improve-machine-learning-workflows/