SoL-Pi: An Agent Harness That Cuts Token Traffic Nearly in Half
Loading video
Loading videoNVIDIA moves recursive self-improvement to the harness layer, scaling auto-research loops across increasingly numerous and diverse environments until four reusable mechanisms survive selection: action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench it performs comparably to Pi across GPT-5.6 Sol and Opus 5 while cutting recorded token traffic by 44.7-49.0% and API cost by about one third, an estimated $8.75-$13.50 saved per hour against native Codex and Claude Code harnesses.
Category: harness
Author: @HuggingPapers
Date: 2026-09-18T00:00:00
Duration: 60.0s
Reference: https://arxiv.org/abs/2609.20519
