Most AI agents are not expensive because they are smart. They are expensive because they keep rereading the same junk. This field note shows how to save LLM tokens with prompt caching, session handoffs, context trimming, and compression tools without pretending every benchmark is gospel...