Interesting, but this doesn't support multiple tools with different arguments, or an arbitrary number of steps in the conversation. If you add that, you basically get PydanticAI.
This can pretty easily be generalized to any number of steps - you're just passing the entire conversation when routing the query. I agree it's more constrained - intentionally. With each step being stochastic, it becomes incredibly challenging to troubleshoot agents when they're doing unconstrained work. Regardless, I look forward to trying this out at some point.