According to @_avichawla, agent ops like routing, guardrails, caching, and evals matter more than core logic for reliable AI deployment.