DataSmith agent beats harnesses tuning architecture, using 200x fewer tokens
jefrankle · x · 2026-08-19
Datology AI introduced DataSmith, an auto agent for data research, arguing data quality is the ultimate compute multiplier.- Through data interventions alone, DataSmith outperforms harnesses that also intervene on architecture and optimizers, beating Claude Code by 5.1 pts on average.- Example: with only 734M post-training tokens — 200x fewer than the default recipe — it improved SmolLM3-3B's post-training performance by 3 pp.- Its design leads it to experiment more, reason more, explore more novel datasets, and spawn more subagents, finding counterintuitive solutions — e.g., determining the best way to boost coding performance was adding more on-policy math data, not more code data.
Related event: DatologyAI Launches DataSmith, an Autonomous Data Research Agent(3 posts)→
More from coding & agent
- Six-month prediction: hiring for generic agent harnesses built from scratch will cool — yi111 · 2026-10-16
- Omarchy ISO wouldn't boot on NVIDIA, so he had Claude patch it — indragie · 2026-10-16
- Dev uses Grok to clean up 3D printer firmware, catches it sneaking in nonsense code — yacineMTB · 2026-10-16
- A rude AI agent took 89% of a shared resource in 102-run multi-agent personality experiment — Brave-History-6502 · 2026-10-16
- Guardrails can make risky delegation work — but it takes a full-time team, says iandanforth — iandanforth · 2026-10-16
- What can you truly delegate? Only tasks with unambiguous, verifiable results — iandanforth · 2026-10-16