Nvidia unveils Sol-Pi token-efficiency mechanisms for AI agents

Nvidia released Sol-Pi, a study of four token-efficiency mechanisms for the agent harness layer, discovered via automated search. In tests, 20 Sol-Pi workers reached better results at 26.8% lower cost than 20 Pi workers.

• Sol-Pi cuts cost by about a third while retaining roughly 94% of quality
• 20 Sol-Pi workers cost 26.8% less than 20 Pi workers
• On Terminal-Bench 4 it solved 15 vs 18 tasks, but cheaper per solve: $14.07 vs $15.91
• Mechanisms: Action Fusion, Online Context Compact, ObservationPack and Evidence-Preserving Reducer