SoL-Pi

SoL-Pi is a research project that optimizes coding agent efficiency through recursive self-improvement. It uses auto-research loops to identify and combine mechanisms that reduce token consumption, cost, and conversation turns without sacrificing work output or evidence quality. Built with Claude Code, SoL-Pi demonstrates cost savings of $4.36–$13.50 per hour compared to existing agent harnesses like Codex and Claude Code. The system constructs executable environments from public data, observes how models explore trajectories, and transforms those observations into efficiency improvements. Designed for researchers and teams running extended agent operations.