I am thinking of something like what co-pilot has as prompts.
We already know based on Kagi benchmarks and knowledge about LLMS what model is better for which sort of task. When Kagi Assistant is opened, a few broad prompts could be made available to the user to know what they are about to use Kagi Assistant for. Example from co-pilot prompts:
Write a short essay that analyzes merits of basic universal income
or
What's good budget hotel chain in Miami?
or
Create motivational song desing for gym workout?
or
My mom's enchilada recipe
or
Check my code for sanity and smell of socks.
This will then in the background select appropriate LLM that is based on speed, accuracy and tokens used vs benefit gained sort of algorithm that's pre-defined. 
What will happen is I click : I am after a recipe and any easy going LLM with Kagi research agent will be sufficient for the task and reasoning models will not be used.
Of course once the query is ongoing, the user might see the usual drop down if they wish to change half way through.
But by initiating on the right LLM users will see value in using the correct tool for the job they are interested in.
