AI CODING AGENTS MISS CRITICAL LINES, STUDY FINDS

The research demonstrates a fundamental limitation in how AI agents approach code fixing. While these models reliably pinpoint the right file in a codebase, they struggle to identify the exact lines that require modification. The SWE-Explore benchmark is the first to separately evaluate code search performance from repair accuracy. This distinction matters: even the most sophisticated fix will fail without sufficient context about which code sections need attention. The findings suggest that AI coding agents need better techniques for narrowing their focus within large files. Current approaches appear to retrieve broad file context but lack precision in targeting critical code segments. These limitations have practical implications for developers relying on AI assistants. Without improvements in line-level accuracy, AI coding agents will continue to produce solutions based on incomplete understanding of the problem space, limiting their effectiveness for complex repairs.