NVIDIA Labs: SoL-Pi automatically reduces the cost of coding agent...

NVIDIA Labs: SoL-Pi automatically reduces the cost of coding agent scaffolding

NVIDIA Labs introduced the open-source project SoL-Pi (arXiv:2609.20519), in which agents themselves selected 4 scaffolding optimization mechanisms out of 152 hypotheses, reducing token consumption by 44.7–49.0% and API costs by approximately one-third without a loss in quality on the EdgeBench benchmark.

Compiled by Sergey KostenchukPublished 2026-09-18Updated 2026-09-18

2026-09-18 Research Anthropic

🤖 NVIDIA Labs: AI itself reduced the cost of coding agent scaffolding

In SoL-Pi, agents tested 152 scaffolding improvement hypotheses — 4 mechanisms survived. The search was conducted on 535 environments, and the final check was performed on the isolated EdgeBench benchmark: Pi quality is maintained, token consumption drops by 44.7–49.0%, and API costs — by approximately one-third.

🌍 Usually, auto-research spends more tokens than it saves. SoL-Pi is an open (MIT) template for cost-effective auto-search for improvements: saving $8.75–13.50 per hour compared to native Codex and Claude Code, and $4.36–5.71 compared to base Pi.

👤 It is installed on top of Pi 0.85.1: pi install git:github.com/NVlabs/SoL-Pi. Action Fusion and ObservationPack are safe, but enable the log reducer only after reading SECURITY.md — it sends their contents to an external model.

Source 1: http://nvlabs.github.io/SoL-Pi/ Source 2: https://arxiv.org/abs/2609.20519

Sources