Agent harnesses are the glue that lets large language models complete complex tasks: they coordinate model calls, tool use, and task execution. Because these systems are often iteratively refined to handle new requirements and failures, they can grow complicated over time.
The arXiv preprint introduces SHarP, a saliency-based pruning approach for agent harnesses. Rather than treating every component as equally necessary, SHarP uses saliency signals to identify which parts of a harness can be removed without breaking task performance.
The paper is a single source, so there are no conflicting findings to compare. As a preprint, its claims have not yet undergone peer review.