We prompt an agent, a tool gets invoked, and a response is generated. But what really happens in between? And when something breaks, how do we trace the failure and debug it effectively?