DS-Lighting Toolkit Standardizes AI Agent Harnesses for Data Science Automation

Large Language Model (LLM) agents are increasingly used to automate data science tasks, but their performance often depends on implicit "harnesses" that manage tasks, execution, and evaluation. This lack of explicit definition makes it hard to reproduce or compare results. A new toolkit, DS-Lighting, addresses this by providing a structured framework for designing these harnesses. It breaks down the harness into four distinct layers: data, workflow, execution, and evaluation, allowing for more transparent and reusable agent designs. Experiments show that this explicit approach significantly boosts reproducibility, comparability, and reliability of data science workflows, while also reducing common system-level failures. The toolkit also integrates open-source benchmarks for standardized evaluation.