Can someone give me an impression of how simplified this set of articles is? I was building out remote MCP architecture at the start of the year and came to a similar conclusion as the author around the stateless nature of LLM API's ("An LLM API is stateless. Every turn, you send the entire conversation as a JSON array and get text back. A harness is the program that builds, maintains, and protects that array using a simple loop.").

However, I just figured that this line of thinking was a consequence of a frankly naive implementation, and that frameworks like LangGraph etc provide a more sophisticated way of thinking about it.