THE ERA OF SKILL GROWING AGENTS:SHIFT FROM TASK EXECUTION TO SKILL GROWTH

Today's agents can browse the web, invoke tools, and execute multi-step reasoning, yet they still operate as if every task were isolated. Their behavior is optimized for task execution, but little capability persists beyond the immediate interaction. The result is increasingly powerful execution without sustained growth. We argue that the next frontier of agent intelligence lies in skill growth: the ability to expand, organize, and effectively operate over an evolving space of executable skills. Rather than focusing solely on how well agents execute a fixed set of tools, we present a structured view that distinguishes crafted execution,experience-guided evolution under limited skills,and regimes in which capability is dominated by large-scale skill access and activation.This perspective clarifies why recent advances increasingly arise from expanding skill inventories,standardized protocols, and execution platforms,and why execution refinement alone tends to saturate.We identify key challenges for skill-growing agents, including operating over open skill spaces, coordinating massive toolsets, and ensuring reliability and efficiency at scale. Overall, this paper calls for a shift from optimizing task execution to designing agents that can grow through skills.