Your agent's thinking and its actions are two different layers. Build a tool-calling agent on WEC Inference, trace it with Langfuse, then debug two real failures from the trace — including the confident, wrong answer that never throws a stack trace.