Self-improving agent harness builds XGBoost from scratch in C for under $0.20
Ordinary-Ad-5639 · reddit · 2026-08-31
- The author stress-tested GPT-5.6 Sol with their self-improving RLM harness Ares, tasking it with building an ML library in pure C, including a full XGBoost implementation.
- Trained 100 trees on Boston Housing: RMSE 68758, MAE 52255, R² 0.8052.
- Only 0.35% off the reference Python XGBoost on the same split and hyperparameters.
- Total cost under $0.20, context usage 1% (15k/1M tokens).
More from coding & agent
- Developer Switches Agent to OpenHands for Better Hackability — morgymcg · 2026-08-31
- Open Source video-use: Edit Videos with Claude Code Agents — Shruti_0810 · 2026-08-31
- Delete and Revalidate Agent Skills Often to Avoid Early Standardization — jdjohnson · 2026-08-31
- Teaching your agent how to use an app or OS is the perfect use for skills — BLUECOW009 · 2026-08-31
- BYOA: Websites Should Offer Skills, Not Agents — thisiskp_ · 2026-08-31
- Frontier coding agents lack self-knowledge, leading to bloated codebases — MinqiJiang · 2026-08-31