As coding agents grow more capable, robust evaluation demands a wider and more diverse supply of verifiable software engineering (SWE) tasks. But real-world tasks are constrained by how often issues naturally occur, which limits the size and variety of available benchmarks.

The paper introduces SWE-at-Scale, a method for synthesizing issues from historical bug fixes. By generating tasks from past fixes, it seeks to overcome the bottleneck of scarce real-world issues and supply evaluation sets that better match the capabilities of modern agents.

According to the abstract, the goal is to enable scalable creation of verifiable SWE tasks, though the full methodology and results are not detailed in the available excerpt.