SoL-Pi: Auto-research loops optimize coding agent harnesses, cutting tokens 44-49%

burny_tech · x · 2026-09-20

A paper from NVIDIA, MIT and collaborators, SoL-Pi applies RSI-inspired auto-research loops at the harness layer, letting coding agents discover their own framework optimizations instead of hand-designing them.

Across diverse environments, the selection process yields four surviving mechanisms—action execution, context compaction, observation handling, and delegated reading—that make up SoL-Pi.

Key results: On the 51-task EdgeBench, SoL-Pi matches Pi's performance with GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7–49% and API cost by roughly one-third—estimated savings of $8.75–13.50/hour vs. native Codex and Claude Code harnesses, and $4.36–5.71 vs. Pi.

Original post →

More from coding & agent

coding & agent channel →