Failure analysis separates model errors from missing context, poor retrieval, ambiguous tools, environment faults, unsafe permissions, and flawed graders. A useful taxonomy turns recurring failures into targeted fixes and new evaluations.