Optimizing Tool Selection for LLM Workflows: Differentiable Programming with PyTorch and DSPy

How local, learnable routers can reduce token overhead, lower costs, and bring structure back to agentic workflows.