Noesion Radarnoesion.ai
Research2026-09-17

NVIDIA researchers introduced SoL-Pi, a harness for recursive agent self-improvement. It scales auto-research loops across many environments. Four mechanisms form SoL-Pi: action execution, context compaction, observation handling, and delegated reading. On EdgeBench, SoL-Pi matches Pi’s performance on GPT-5.6 Sol and Opus 5. It cuts token traffic by 44.7-49.0%. API costs drop by about one third. Estimated hourly savings reach $8.75-$13.50 versus Codex and Claude Code. Savings are $4.36-$5.71 versus Pi. The paper and code are publicly available.

What changed

Coding agents used isolated predictions with higher token and API costs.

What it unlocks

Running long, recursive self-improvement loops at lower cost and token use.

  • 44.7-49.0% token traffic reduction
  • $8.75-$13.50 hourly savings vs Codex/Claude Code
  • $4.36-$5.71 hourly savings vs Pi
  • 51-task EdgeBench evaluation

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.