Scaling Laws of Skills in LLM Agent Systems Identified

A recent study on arXiv identifies two interconnected scaling laws that influence skill libraries within LLM agent systems. Researchers examined 15 advanced LLMs, 1,141 practical skills, and more than 3 million routing or execution choices. Their findings indicate that the accuracy of single-step routing decreases logarithmically as the library expands (R²>0.97 across all models). Errors transition from local skill competition to broader cross-family drift, often leading to the dominance of overly general 'black-hole skills.' The execution law suggests that prior to state realization, joint routing behaves in a nearly multiplicative manner, while correct execution can enhance challenging downstream decisions by roughly 4×. A single parameter, the routing logarithmic decay slope b, links the two laws, allowing routing-side adjustments to forecast execution-side improvements across various models. These laws offer practical insights for enhancing agent systems.