There is no single shape of work
One job may include a quick classification, a long-context synthesis, a code change, and a polished piece of writing. Treating all of those steps as the same workload leaves quality, speed, or cost on the table.
The application has more context about each step than the person making the request. It knows whether the task needs tools, how much context it carries, and what kind of output the next step expects.
Routing belongs below the interface
A model picker asks the user to make an infrastructure decision before they can describe the outcome. That choice also goes stale as models and providers change.
Routing lets the product select a model for the work in front of it and change that selection without rewriting the user’s workflow. The request remains stable even when the underlying model does not.
Keep the tradeoffs visible
Abstraction should not mean mystery. Teams still need to understand what happened, inspect sources, and apply policy to the providers their data can reach.
Model-agnostic infrastructure is useful when it creates flexibility without hiding accountability. The system owns the selection; the customer keeps control of the boundary.
One line to keep
“The product should absorb model complexity so the customer can stay focused on the job.”
Published August 7, 2026
Back to the blog