Scaling Coding Agents via Atomic Skills

Source: arXiv Authors: Yingwei Ma, Yue Liu, Xinlong Yang, Yanhao Li, Kelin Fu, Yibo Miao, Yuchong Xie, Zhexu Wang, Shing-Chi Cheung Date: 2026-04-06 Primary category: cs.SE All categories: cs.SE, cs.AI

Abstract

This paper argues that coding agents should scale by mastering composable atomic skills such as localization, editing, test generation, issue reproduction, and code review. Improvements on these atomic skills transfer to broader software tasks better than direct optimization on composite benchmarks alone. It matters here because a self-evolving workflow system may need to evolve both workflow graphs and the reusable skill atoms those graphs invoke.