Not Every Node in Your Agent Needs an LLM

A design pattern for building agent pipelines you can actually reason about Look at the last agent diagram you drew. Count the boxes. Now count the boxes that call an LLM. If those numbers match, you have a problem. No amount of better prompting will fix it. The habit is everywhere. Classification - LLM. Routing - LLM. Retrieval ranking - LLM. Validation - LLM-as-judge. Repair - LLM. Every node is a model call. The whole pipeline collapses into one shape: a chain of probabilistic steps, each capable of failing in its own way, with no deterministic ground anywhere. It feels powerful.