If you write LLM calls in a production environment, you often hardcode the model name directly in the code. Each time you need to change it, you have to manually modify the code and test it.

You can try running it — it provides a runtime policy layer. Instead of locking a specific model in the code, it allows each request to carry a policy.
This policy automatically filters out models that don’t meet the rules at runtime and sorts the remaining eligible models, for example, by price. If the selected model fails, it can automatically fall back to the next eligible option.
It’s like passing a temperature parameter in an API call — routing decisions are no longer hardcoded in the code but become part of the request, dynamically determined at runtime by the policy. Compared to hardcoding models in the code or doing static configuration on a gateway dashboard, this approach offers more flexibility and reduces ongoing manual maintenance.