A benchmarking study by Brandon Foley has produced one of the cleaner empirical assessments of where AI coding agents succeed and where they fail at real-world software engineering. Foley ran three AI agent configurations against nine open pull requests from the Kubernetes repository and found that while the agents reliably located and patched isolated defects, they consistently failed to understand system-wide implications. The result challenges the prevailing assumption that better code retrieval is the primary lever for improving automated bug fixing.