Abstract
LLM agents often mis-call tools, and it is commonly assumed that the model fails to see the right tool in a crowded harness. However, we investigate this through an overlooked lens: the model's attention to labeled tool-definition segments. In analyzing real BFCL failures, we find that by per-candidate attention argmax, the model attends to the correct tool 80% of the time (compared to a 21% chance), while the gold standard is under-attended only 10% of the time. This indicates that the model can see the right tool but still makes incorrect selections. This directly refutes the intuitive 'crowded-harness/lost-in-the-middle' explanation: the failure occurs at the decision readout, not in the harness. We confirm this in three ways: (1) Input vs. readout: repairing the prompt (reordering or duplicating the gold tool) recovers the model's selection ability.