Title: Addressing 'Skill Issues': Enhancing Lakehouse Agents Through Data-Centric Optimization

Abstract: As coding agents increasingly interact with data infrastructure, their effectiveness is determined not merely by the quality of the underlying models, but also by the proficiency and environmental configurations that instruct them on system usage. This study investigates the optimization of these specific artifacts for agents functioning within a branching lakehouse environment known as Bauplan. In this framework, headless APIs alongside Git-like data primitives facilitate data workflows via code, utilizing branches, commits, and merges. Our primary finding is that a branching lakehouse transforms the evaluation of data agents from a matter of matching outputs to one of verifying states, as the pipeline code generated by agents produces tangible, inspectable modifications within the lakehouse. We introduce a data-centric optimization pipeline designed to create task-verifier pairs, run candidate skills in isolated sandboxes, and evaluate trajectories through both trace-level indicators and programmatic assessments of the lakehouse state. Preliminary testing across 25 tasks demonstrated that optimized skills increased accuracy by 31.9%. These findings indicate that data workflows involving write operations offer a valuable foundation for refining agent capabilities, extending beyond the scope of read-only tasks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC