Nvidia's SoL-Pi System Cuts Coding Agent Token Usage Nearly in Half
Nvidia researchers have created a system called SoL-Pi that optimizes the harness, the control layer between the model and its environment, to cut costs. The harness is responsible for how an agent sees states, runs actions, and processes feedback.
SoL-Pi automates the process of optimizing the harness by using a research AI that analyzes agent traces, proposes changes, and tests them in prepared environments. This approach draws on recursive self-improvement.
The system was tested across 535 executable environments, exploring 152 directions and generating over 3,000 runs and 60,000 agent-environment interactions. The results showed a 44.7 to 49 percent reduction in token usage compared to the original Pi harness, with some configurations reaching up to 93.7 percent of Pi's score.
The efficiency gains come with trade-offs, as shorter context can reduce prompt cache reuse and lead to increased costs. However, Nvidia suggests that pretraining the harness across many tasks could make searching for its successor cheaper.