The LLM tax on every orchestration call

Martin Fowler is writing about something most people have felt but not named: every time a chatbot layer sits between the user and a real tool — a search engine, a database, a code runner — the LLM has to decide whether the user's words are ready to pass through. That decision costs tokens, and tokens cost money. The more layers you stack, the more you pay.

The real cost isn't the LLM call itself. It's the calls it triggers. An orchestrator that looks up a definition, then queries a database, then formats a response will fire three LLM calls for what a human would do with one. Each one burns tokens. Each one adds latency. The bill compounds quietly.

Why this matters for us: every app the primos are building on top of LLMs is quietly paying this tax — and most of them don't know it yet, so they're pricing wrong.