Notes on LLM context, where noise builds up, and what to do about it.
I tested both ways of reading an n8n workflow in production: straight through MCP, and through a deterministic reducer. One hit a token ceiling error. The other cut 84%.