SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang and 6 others
NVIDIA · NTU · MIT
Published September 17, 2026 · on the feed September 18, 2026 · ▲ 130 · arXiv 2609.20519
The abstract
As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback. Token efficiency therefore becomes important for scaling recursive self-improvement. We take an RSI-inspired approach at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts. At this scale, the process yields reusable improvements that transfer beyond their development setting, moving automated harness discovery toward production-level outcomes. Four mechanisms survive selection and form SoL-Pi, spanning action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench evaluation, SoL-Pi achieves performance comparable to Pi across GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7-49.0% and API cost by about one third. In other words, estimated hourly savings are \8.75-13.50 relative to native Codex and Claude Code harnesses, and \4.36-5.71 relative to Pi.
The Brief
2:00A short podcast summary of the paper. Written and voiced by AI.
Transcript
SoL-Pi. Recursively Scaling Auto-Research Loops for Efficient Agent Harness. By Haozhe Liu and thirteen others, from NVIDIA, N T U, and one other institution. Published on the seventeenth of September. Its code repository is worked out: about forty-five source files, with a readme and a licence.
Coding agents are moving from supervised code completion to exploration that runs unattended, around the clock. Their work grows from isolated predictions into long trajectories of reasoning, tool use and feedback. At that scale, token efficiency becomes important for scaling recursive self-improvement. When an agent runs all day, every unnecessary token is paid for again and again.
The approach takes its inspiration from recursive self-improvement, and applies it at the harness layer. Auto-research loops are scaled across increasingly numerous and diverse environments for harness rollouts. At that scale, the authors report, the process yields reusable improvements that transfer beyond the setting where they were found.
Four mechanisms survive selection and together form SoL-Pi. They span action execution, context compaction, handling of observations, and delegated reading. Survival is the operative word. The loops produce candidate improvements, and these four are the ones that held up under selection.
On the fifty-one task EdgeBench evaluation, SoL-Pi performs comparably to Pi, across GPT-5.6 Sol and Opus 5, while cutting recorded token traffic by forty-four point seven to forty-nine percent, and API cost by about one third. Estimated hourly savings are eight point seven five to thirteen point five dollars relative to the native Codex and Claude Code harnesses, and four point three six to five point seven one dollars relative to Pi. Put another way, the same work gets done at about two thirds of the API cost.
The result moves automated harness discovery toward production-level outcomes. The harness improvements were not designed by hand. They were what remained after many automated loops, and what remained is a harness that does the same work while reading and writing far fewer tokens.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is an agent that runs all day and bills you for every token twice over.
You run a coding agent that works unattended, around the clock. Over a long job it produces long trajectories of reasoning, tool calls and feedback, and you pay for every token it reads and writes.
The road ahead
- 1 · The problem ~6 min An agent that works all day re-reads and re-writes so many tokens that waste, paid for again and again, becomes the main cost.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min An AI optimiser studies agent traces, proposes harness changes, tests them across hundreds of environments, and keeps the few that hold up on held-out tasks.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Broad auto-research loops propose and test harness changes; those that transfer are kept, four survive, they are combined into SoL-Pi, and EdgeBench checks it.
- 6 · The experiment ~6 min You’ll guess: SoL-Pi was tuned to spend far fewer tokens than the Pi harness it was built from. What happened to its task performance on EdgeBench?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.