NVIDIA open-sources SoL-Pi: auto-research loops cut agent token cost by up to 54%

udmrzn · x · 2026-09-11

NVIDIA's NVLabs open-sourced SoL-Pi (MIT license), a Pi-based agent-harness efficiency layer discovered via scaled auto-research loops. It builds 535 executable training environments (495 from real GitHub Issue-PRs, 40 synthetic with verifiers) and lets an auto-research loop propose, implement, and validate harness improvements — only 1/40 of ideas survive.

Four winning optimizations: Action Fusion (merge test/run into one tool call), Online Context Compact (compact at subtask completion points), ObservationPack (index oversized tool results instead of replaying them), and Evidence-Preserving Reducer (cheap agent compresses long logs while cited evidence is verified item by item).

Results: vs original Pi, 45–49% fewer tokens and 1/3 lower cost at 94% of average task score; vs native harnesses, 35–64% fewer tokens and 50–54% lower API-priced cost. On GPT-5.6 Sol, SoL-Pi even beats the native Codex harness on EdgeBench. The author notes this converges with DeepSeek's harness direction.

Related event: NVIDIA Open-Sources SoL-Pi, Cutting Pi Coding Agent Token Costs(2 posts)→

Original post →

More from coding & agent

coding & agent channel →