Coding agents that answer analytical questions over a data lake typically discover, clean, and join relevant data for each query. According to a new arXiv paper, they then discard all that work and start from scratch on the next question. The authors describe this as a key inefficiency in current approaches.
The paper introduces "tool evolution," an unsupervised method in which agents turn the steps they have already performed into reusable tools. Instead of repeating the same data preparation and joining logic, future queries can draw on these evolved tools. The abstract does not provide implementation details, but the framing suggests a move toward cumulative learning in data-lake analytics.
The significance is practical: analytical workloads over large, heterogeneous data lakes are expensive, and much of the effort is repetitive. If agents can retain and generalize what they have learned, they may answer subsequent questions faster and with less wasted computation. The paper is a single source, so no alternative views are presented here.