DS-Lighting Toolkit Decouples Data Science Agent Harness Layers

DS-Lighting introduces a modular toolkit that decomposes agent harnesses into four layers: data, workflow, execution, and evaluation. It represents agents as executable operator programs to enable reproducible comparisons across heterogeneous data science tasks using an MLE-Bench style format.