SoL-Pi

SoL-Pi is a set of coding-agent harness optimizations introduced by Haozhe Liu and coauthors from NVIDIA, Nanyang Technological University, and MIT in "SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness" (arXiv:2609.20519, September 2026). Instead of hand-designing how a coding agent's harness manages tool calls, context, and observations, the authors run an "RSI-inspired" auto-research loop: many candidate harness modifications are tried across a large and diverse set of coding environments, and only the modifications that keep improving results across that whole range, rather than just one setting, survive.

Four mechanisms survived this process and make up SoL-Pi. They span action execution, context compaction, observation handling, and delegated reading, which offloads part of the work of reading and summarizing tool output to a separate step. None of them change the underlying language model's weights. They change only the harness code that decides what the agent reads, remembers, and acts on between steps.

On the 51-task EdgeBench evaluation, SoL-Pi matches the task performance of an existing harness called Pi across two different underlying models, GPT-5.6 Sol and Opus 5, while reducing recorded token traffic by 44.7 to 49.0% and API cost by about one third. The authors estimate this saves roughly $8.75 to $13.50 per hour compared with running the same models through the native Codex or Claude Code harnesses, and $4.36 to $5.71 per hour compared with Pi.