Traditional benchmarks have shortcomings for tool-calling agents: they are typically static question-answer pairs and cannot simulate real-world side effects and state changes.