Designing for the Day Your Model Gets Deprecated
A thought experiment on treating a specific model version as a dependency that will eventually be retired, and building the system so that day is routine instead of a crisis.
This one's speculative — a thought experiment about design principles, not a report on something I've built.
Every model version is a dependency with an expiration date
A specific model version is a dependency exactly like a software library, except with a shorter and less predictable lifespan — providers retire old model versions on their own schedule, and a system built as if the current model will be available forever is quietly accumulating the same kind of risk as pinning a critical library to a version nobody plans to ever update.
Prompts tuned to one model's quirks are the hidden coupling
The dependency isn't just the model name in a config file — it's every prompt that was iterated against one specific model's particular tendencies, formatting habits, and failure modes. Swap the model out and a prompt that worked well can silently produce subtly different behavior, even with functionally similar underlying capability, because it was unconsciously tuned to compensate for quirks the new model doesn't have and doesn't need compensating for.
Build the regression suite before you need it
The same behavioral regression tests that catch a bad prompt edit are exactly what makes a model migration tractable instead of terrifying — run the existing suite against a candidate replacement model before switching, and the gap between "seems fine in a quick manual check" and "actually verified against real expected behavior" closes considerably.
Plan the obituary before you need it
The healthiest version of this discipline treats model deprecation as a known, recurring maintenance event on a calendar, not a fire drill that starts the day a provider sends a deprecation notice with a deadline attached. A system that's been quietly tested against next-generation models all along has an easy migration. A system discovering its dependency for the first time on a deadline does not.
I'm Jesse Myers — Marine veteran, 32 years in enterprise IT, now building production AI systems. This site is where I write about what I've actually built, and occasionally about ideas I haven't built yet but think are worth taking seriously.