Cutting LLM token Usage by ~80% using REPL driven document analysis

How much context do I have to use on every message so that the agent knows how to use this bespoke language and repl?I can achieve the same thing with traditional tool calls and a state map attached to a session that I can selectively expose to the agent. I actually do this already with ADK and the article has given me some ideas for sure, concept cool, but...Why did the author choose to do this with a new language and custom repl?