Submitted to arXiv on September 17, 2026 (paper 2609.20519) by Haozhe Liu, Enze Xie, Song Han and eleven co-authors, SoL-Pi applies an idea borrowed from recursive self-improvement to the least glamorous layer of a coding agent: the harness, the code that decides how the model executes actions, what it sees and what it keeps in context. The authors argue that as coding agents move from completing code to running unattended around the clock, token efficiency becomes a scaling constraint in its own right.
Rather than hand-tuning the harness, the team ran automated research loops that propose harness changes and test them across increasingly numerous and diverse environments. At enough scale, they report, the process yields improvements that transfer beyond the setting they were developed in. Four mechanisms survived selection and make up SoL-Pi: changes to action execution, context compaction, observation handling, and delegated reading.
On the 51-task EdgeBench evaluation, SoL-Pi performed comparably to the Pi harness it builds on, with both GPT-5.6 Sol and Opus 5 as the underlying model, while reducing recorded token traffic by 44.7 to 49.0 percent and API cost by about one third. The authors translate this into estimated savings of $8.75 to $13.50 per hour against the native Codex and Claude Code harnesses, and $4.36 to $5.71 per hour against Pi.
The claim is about efficiency at equal performance, not a capability gain, and it rests on a single 51-task benchmark. Hourly savings figures depend on assumptions about usage and pricing that will not match every deployment. Its broader interest is as another data point, alongside work like StateM, that the harness around a model is a large and still under-optimized lever - and that agents can now do some of that optimization themselves.