Agents shift slop to design decisions

An analysis of frontier AI coding models reveals that while line-level code quality and solve rates continue to improve, overall slop has moved up the stack to infest design decisions, error handling, and library reuse. Benchmarks such as SlopCodeBench, LibraryDesignBench, and LibraryUseBench demonstrate that agents frequently fail to reuse existing library code and instead reimplement functionality with poor abstractions. The author argues that future evaluations must incorporate longer-horizon benchmarks that account for how agents interact with code produced by other agents. Open article →