Model providers update APIs, release new versions, deprecate old ones, and adjust pricing on a timescale of months, sometimes weeks. Agent logic changes more slowly than that.
Production software already solved the same problem for storage. The application does not know which database host it is talking to. It talks to an abstraction layer that routes to the appropriate host, because the host changes more often than the logic built on top of it, and coupling the two creates a migration cost that compounds with every change. David Parnas named the principle information hiding in 1972: isolate the decision most likely to change behind an interface that does not change with it.
Among the AI agent architecture patterns I have shipped in production, the one I return to most applies that idea to model selection, and enterprises skip it more than any other.…