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: general coding agents beat specialized data agents by 37 points(4 posts)→
More from coding & agent
- WanGP adds one-click AMD GPU support for Tencent's Hy-Nomas world model, runs on 8GB VRAM — cocktailpeanut · 2026-10-27
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Founder generates 46 app promo videos in 4 days with Claude, skipping $11,500 in UGC fees — PrajwalTomar_ · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27