NVIDIA unveils SoL-Pi: auto-research loops cut coding agent token traffic by up to 49%

NVIDIA, NTU and MIT researchers released SoL-Pi, a set of four efficiency mechanisms for the open-source Pi coding agent, discovered through auto-research loops at the harness layer. On the 51-task EdgeBench, it cuts token traffic by 44.7–49.0% and API cost by about 33% while keeping roughly 94% of Pi's score. The MIT-licensed extension ships on GitHub under NVlabs and runs on an unmodified Pi 0.85.1.

• Cuts EdgeBench token traffic by 44.7–49.0% versus Pi
• Lowers API cost about 33% while keeping ~94% of Pi's score
• Search covered 152 directions, 535 environments and 3,000+ runs
• MIT-licensed extension on GitHub, tested with Pi 0.85.1 and Node.js 22.19+