Agentic vs. Vector Search: An Eval-Driven Approach to Coding Agent Performance