RealSWE Benchmark Exposes Gap Between Coding Agents and Real Users

Analysis shows 88% of real user requests lack the structured context found in SWE-bench, which contains only 7% of such cases. RealSWE introduces 381 task families to better evaluate agents against casual, short, and unstructured prompts.