A counterintuitive bit of product advice for agentic software: use the worst model you can get away with.
If you do, then every model that is better will definitely work.
You can more easily switch between models, and potentially save a lot of money.
If you see what you can do with the best models, you build systems that require the best models.
But if you make a system that is resilient to low quality, you can make robust, cheap solutions.
Kind of like trading opex for capex.
A little bit of cleverness in your structure up front saves you ongoing opex on more expensive models.
The larger you scale, the more this saves.