One thing I'm proudest of in my upcoming course is that we use a request logger to read the raw stuff being sent to Anthropic/OpenAI This has a bunch of cool effects: - Makes all the mechanics completely clear: sessions, context window, model/harness boundary - Instills context paranoia - every token you send matters, and it's easy to see waste - Empowers folks to want to build their own harnesses - all the magic disappears