Tool calls are where conversation becomes action — and where hallucinated arguments become real-world mistakes. Testing tool-call accuracy is testing what the agent actually does.