Nvidia's SoL-Pi cuts coding agent token usage by 50% with optimized harness

nvidia's sol-pi system targets the harness, the control layer between the model and environment used by coding agents like codex and claude code. sol-pi uses a research ai to analyze agent traces, propose harness changes, and test them in prepared environments. the system explored 152 directions across 535 executable environments, generating over 3,000 runs and 60,000 agent-environment interactions. sol-pi produced four mechanisms to eliminate wasted work: action fusion, online context compact, observationpack, and the evidence-preserving reducer. these changes resulted in a 50% reduction in token usage compared to codex and 54.3% compared to claude code on edgebench, without sacrificing performance.