NVIDIA's SoL-Pi auto-evolves agent harnesses, cutting tokens ~50% and API costs ~33% with no quality loss

omarsar0 · x · 2026-09-19

An NVIDIA paper on self-evolving agent harnesses introduces SoL-Pi: instead of hand-tuning, it runs auto-research loops at the harness layer across repository-derived and verifier-driven environments, keeping only mechanisms that survive selection. Four survived: Action Fusion (changes how actions execute), Online Context Compact (mid-run compaction), ObservationPack (reshapes observation handling), and Evidence-Preserving Reducer (delegated reading). On a 51-task EdgeBench eval, SoL-Pi cuts token traffic by nearly half and API costs by about a third ($8.75–$13.50/hour vs native Codex and Claude Code harnesses) while matching baseline harness performance on GPT-5.6 Sol and Opus 5. Takeaway: build your own harness and let evolution do the tuning.

Related event: NVIDIA's SoL-Pi: Self-Evolving Agent Harness Cuts Token Use by 45%(3 posts)→

Original post →

More from coding & agent

coding & agent channel →