Coding agents beat hand-built data agents by 37 points with 4x fewer turns, paper finds
CShorten30 · x · 2026-09-08
A paper by Liana Patel, Matei Zaharia, Ion Stoica and others asks what systems research remains for data agents under the Bitter Lesson.
- Tracking two years of data agent evolution, they find general-purpose coding agents beat hand-designed data agents by up to 37 points with 4x fewer turns.
- They argue system layers built to compensate for model limitations will increasingly be absorbed by the models themselves.
- Enduring opportunities lie in persistent semantic context: curated contextual information about the data environment served across many queries.
- Future data systems should natively serve persistent semantic context as a first-class abstraction, opening research into efficient context data structures, storage, compression, and semantic consistency protocols for integrity over huge knowledge corpora.
Related event: Study: general coding agents beat specialized data agents by 37 points(4 posts)→
More from coding & agent
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26