← All writing

The model is the most replaceable part of your stack

August 2026 · 6 min read · vallesAI

Every few months, the ranking of frontier models reshuffles. A provider ships something better, or dramatically cheaper, or both; another deprecates the model you built on with twelve months' notice if you're lucky. This is not a temporary condition of an immature market. It is what a healthy, competitive model market looks like, and there is no reason to expect it to stop.

Most AI systems we're asked to look at were built as if this weren't true. The model isn't a component in these systems — it's a load-bearing wall. Prompts assume one provider's formatting quirks. Tool-calling code is written against one API's schema. Retry logic, token accounting, and cost assumptions are scattered through the codebase, each quietly hard-coding the vendor. When a better option ships, the honest cost of switching is a rewrite, so nobody switches, and the system gets slower and more expensive relative to the frontier every quarter it stands still.

Lock-in is rarely a contract you signed. It's an architecture you drifted into.

What model-agnostic actually means

It does not mean using some abstraction layer that promises to make every model interchangeable — models differ in real ways, and pretending otherwise costs you their strengths. It means three concrete engineering decisions:

• The model sits behind an interface you own. One place in the codebase knows which provider is being called. Everything else — pipelines, tools, UI — talks to your interface, not theirs.
• Prompts are assets, not string literals. Versioned, tested, and organized so that adapting them to a different model is a bounded task with a diff, not an archaeology project.
• An evaluation harness stands between you and any change. Switching models — or upgrading within a provider — is a measured decision: run the candidate against your evaluation set, compare accuracy, cost, and latency, then decide.

The harness is the point

Of the three, the evaluation harness matters most, because it converts the market's churn from a threat into a dividend. When a new model ships, a team with a harness spends an afternoon and gets a number: switching saves 40% at equal accuracy, or doesn't. A team without one gets a meeting, then another meeting, then the status quo.

We've seen the same pattern enough times to state it as a rule: the durable assets in an AI system are the data pipelines, the evaluation suite, and the institutional definition of what "good" means for the problem. The model is the most replaceable part — and the systems that treat it that way are the ones still improving two years after launch.