Nvidia researchers have developed SoL-Pi, a system that optimizes the “harness” controlling how a coding agent plans, uses tools and processes results. On the EdgeBench software benchmark, the resulting configurations used 44.7% to 49% fewer tokens while keeping performance near the original Pi agent.
A research agent examined execution traces, proposed harness changes and tested them in prepared environments. The surviving techniques fused consecutive actions, compacted context after planning, archived large tool outputs behind short summaries and reduced error logs with a cheaper model while preserving important evidence. The most efficient combination reached 93.7% of the baseline score; the strongest single technique slightly exceeded the baseline while still reducing token use.
The gains did not transfer perfectly. A harness optimized with GPT-5.6 Sol triggered its mechanisms less often with Opus 5, and SoL-Pi solved 15 of 63 Terminal-Bench tasks compared with 18 for Codex and Pi, although at lower cost. Context compression can also discard instructions or reduce prompt-cache benefits. The research suggests agent software can be optimized as aggressively as models, but savings measured on one task set should not be assumed to hold across every workflow.