The LLM App Isn't a Model, It's a System: Designing for Quarterly Model Swaps

[garybake]

As LLM providers ship breaking changes on a near-monthly cadence, the difference between a 10-minute .env swap and a 400-line PR comes down to architecture. This post introduces seam-driven design for Python LLM apps - five narrow interfaces (provider, prompt, tools, config, observability) that can each be replaced independently - illustrated with a working FastAPI + LangChain reference app. Readers get a concrete checklist and two hands-on drills they can run against the open-source code immediately.