Prioritize routing over model choice to slash AI costs

The article argues that teams building AI agents make a critical mistake by choosing the model first and the architecture second. It advocates for prioritizing a "router" — a small piece of code that determines which tier of model handles each request. When the router is designed correctly, 70-80% of traffic can run on cheap local or async models, dramatically reducing AI costs. It cites Brian Armstrong's example of Coinbase cutting AI spend in half while token usage grew, reinforcing the principle that most AI work doesn't need expensive frontier models.