Stanford paper: general coding agents beat hand-built data agents by up to 37 points

CShorten30 · x · 2026-09-08

A new arXiv paper by Liana Patel, Ion Stoica, Matei Zaharia and colleagues tracks two years of frontier models on data-agent benchmarks: general coding agents now beat carefully hand-designed data agents by up to 37 points with 4× fewer turns. Arguing the Bitter Lesson is subsuming data-system engineering layers, the authors identify enduring research in 'persistent semantic context' — curated context about the data environment served as a first-class abstraction — and outline open problems in context data structures, storage, compression, and semantic consistency protocols.

Related event: Study Finds General Coding Agents Beat Specialized Data Agents by 37 Points(3 posts)→

Original post →

More from coding & agent

coding & agent channel →