A benchmarking study on the CNCF blog tested three AI coding agent configurations against nine real Kubernetes bug reports. The agents differed only in code retrieval strategy: RAG-only (via KAITO RAG Engine + Qdrant), hybrid RAG + local filesystem, and local clone only. All ran Claude Opus 4.6 with the same timeout and output format. RAG-only was fastest (~76s) and cheapest; hybrid was slowest and most expensive due to high model invocation counts. On correctness, the dominant failure was incomplete fixes — agents addressed the immediate bug but missed adjacent changes and system-wide impacts. Retrieval strategy influenced code discovery but not reasoning quality. The strongest predictor of agent success was the quality of the bug report itself: well-specified issues with exact file, function, and expected behavior caused all three approaches to converge on high scores, outweighing retrieval architecture differences.
