AI coding agents find the right file but miss the exact lines that matter, study shows

Researchers have developed a benchmark called SWE-Explore that reveals a specific limitation in current AI coding tools: while these agents can accurately locate the correct file for a task, they frequently fail to identify the precise lines of code that require modification. By isolating code search from the actual repair process, the study demonstrates that these agents struggle with contextual accuracy, which often leads to failed software patches.