Nvidia Adds an External Boost to Pi: Agent Researches on Its Own to Find 4 Optimizations, Saving Up to 64% on Tokens

Beating AI News Flash: NVIDIA NVLabs has open-sourced SoL-Pi, adding a layer of efficiency optimization to the Coding Agent Harness "Pi." The research team built an automated research pipeline that lets multiple Agents analyze Pi's execution traces, propose optimization plans, modify code, and automatically verify changes. There were initially 152 candidate directions, and only 4 were ultimately retained.

All 4 optimizations focus on reducing redundant overhead. For example, after code modifications, tests run directly without an extra model call; large files and long outputs that have been read are cached locally and retrieved when needed; after a subtask ends, it determines whether to compress the context; and overly long logs are first sent to a cheaper model to filter key points, then verified to confirm the content indeed comes from the original logs.

On EdgeBench, SoL-Pi uses 45%–49% fewer Tokens than the original Pi, with costs about one-third lower; compared with the native Harnesses of Codex and Claude Code, it uses 35%–64% fewer Tokens, with costs 50%–54% lower based on API list prices.

However, saving money comes at a slight cost in capability. SoL-Pi retains about 94% of Pi's average score; among the 63 CPU tasks in Terminal-Bench 4, it solved 15, while both Codex and Pi solved 18.