Structured Output vs Function Calling

Current Approach of using the Just Structured Output Feature of LLM (Constrained Decoding) though reliable, is expensive in term of token usage and also in time consumption.

Thats why the Function Calling is widely used in the Agent Frameworks for its fast response.

What if we use a hybrid approach where if the user wants speed, we will use prompt based or function calling approach to next the next decision or another approach where we provide each tool as seperate tools + step transition as a tool. and if it fails fallback to the structured output approach.

Such that we are not giving up the reliability but at the same time gains performance.