Nvidia researchers have introduced SoL-Pi, an automated system that optimizes the harness layer controlling AI coding agents. By targeting the software interface that manages model states, tool calls, and execution feedback, SoL-Pi reduces token usage by up to 49 percent while maintaining 93.7 percent of baseline task performance on the EdgeBench benchmark.

Optimizing agent control layers is notoriously difficult because tool execution, context management, and abort logic are tightly coupled. SoL-Pi automates this process using a secondary research agent that reviews execution traces, proposes harness modifications, and evaluates candidate changes across isolated test environments.

The system discovered four key optimization mechanisms: combining consecutive action steps into single calls, trimming accumulated context after planning phases, summarizing long tool outputs, and using smaller secondary models to compress verbose error logs. Evaluation was kept strictly isolated from the search process to prevent harness overfitting.

Why it matters

  • Harness-level optimization offers a major token cost reduction vector independent of model architecture changes or serving infrastructure.

  • Automating harness tuning reduces operational overhead for enterprise coding agents running long unsupervised execution chains.

Source: the-decoder.com