The LLM is becoming interchangeable. The control architecture is not.
On what actually accumulates value when the execution layer can be replaced in an afternoon.
Substitution is the base case now
The practical question is no longer whether a given model can be swapped for another. Increasingly it can, and the interval between a capability appearing in one provider and appearing in the others keeps shortening. Treating the current choice as permanent is a bet against the clearest trend in the field.
That has an uncomfortable implication for anything built tightly against one provider. Work invested in a specific model's quirks — the phrasing it responds to, its particular tool-calling dialect, the workarounds for its failure modes — is work with a short half-life. It is not wasted, exactly. It simply does not compound.
What compounds instead
The things that hold their value are the ones describing the operation rather than the executor. What the workflow is for. Which decisions require a person. What the permissions actually are, resolved to specifics. Where exceptions go. What evidence a run must leave. How performance is measured, and what a bad number means.
None of that changes when the model changes. All of it is expensive to work out and cheap to keep. It is also the part that cannot be bought, because it is a description of one organisation's operation rather than a general capability.
This is the concrete meaning of model independence. Not neutrality for its own sake, and not an abstract preference for portability — a claim about where durable work should be directed when one layer is depreciating quickly and the other is not.
The test for whether you have it
Ask what would have to change to move a workflow to a different execution provider. If the answer includes the approval rules, the permission model, the escalation paths or the definition of the outcome, those things are entangled with the executor and will be re-litigated at every switch.
If the answer is an adapter and a re-run of the evaluation set, the architecture is independent in the way that matters. The operating definition sits outside the thing doing the work, which is the arrangement that lets the thing doing the work be replaced without renegotiating what it is allowed to do.