LLM guardrails filter model output. Kastra intercepts actions. Different layers, different threat models — this comparison explains where each fits.