Automated harness search cuts agent token use by up to 49% on EdgeBench

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to install: this is a preprint, and no released harness or code is mentioned. The transferable claim is that harness changes can be searched automatically and still hold up in environments other than the ones they were found in. If you pay for long agent rollouts, token traffic, not per-token price, is the line to watch.

A preprint reports that SoL-Pi, an agent harness found by automated search (the scaffolding that supplies a model's context and runs its tools), performs comparably to Pi on EdgeBench while using 44.7% to 49.0% fewer tokens and roughly a third less API spend. EdgeBench has 51 tasks, and the runs cover GPT-5.6 Sol and Opus 5. The authors say recursive search across many agent environments produced four reusable changes, governing action execution, context compaction, observation handling and delegated reading. Estimated hourly savings are 8.75–13.50 against native Codex and Claude Code harnesses, and 4.36–5.71 against Pi. The scores are reported without an evaluation date.

Models in this story

Key facts

  • ·SoL-Pi reduced recorded token traffic by 44.7% to 49.0% while performing comparably to Pi on EdgeBench, which contains 51 tasks source
  • ·API cost fell by about one third; the runs used GPT-5.6 Sol and Opus 5 source
  • ·Estimated hourly savings are 8.75–13.50 against native Codex and Claude Code harnesses and 4.36–5.71 against Pi, with the currency unit not specified in the abstract source
  • ·Four mechanisms survived selection: action execution, context compaction, observation handling and delegated reading source
  • ·The EdgeBench scores are reported without an evaluation date source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire