Hannes Hapke explains how to move beyond post-hoc narratives by using Sparse Autoencoders with Kiji Inspector to uncover true causal mechanistic interpretability in production LLM agent decisions.