Why AI Agents Excel at Starting Software But Struggle to Finish It

AI coding agents can rapidly generate the first 80% of a software project, but completing the final stretch remains a largely unsolved challenge for developers. The bottleneck has shifted from building code to hardening it — handling edge cases, integration issues, and failure modes that require careful human judgment. At KandiForge, developer and author experienced this firsthand when agents produced millions of lines of code across 20 repositories, yet finishing the work proved unmanageable at that scale. Adding more reviewers and testers helped only partially, as tests were typically written after code was already integrated, meaning problems were caught too late. The author is documenting ongoing experiments to address this 'last mile' problem in a forthcoming series, ahead of a talk at the BuildStuff conference in Vilnius in December.