Trace

Why it matters

In traditional software, a stack trace tells you exactly which function threw an error. In AI applications, the equivalent question is harder because failures are often semantic rather than structural. The model returned a response, but was it the right response? Traces capture the full execution path of a request, including every LLM call, tool invocation, retrieval step, and intermediate output, so you can reconstruct what happened and why. Without traces, debugging an AI application means guessing which part of a multi-step pipeline went wrong. With traces, you can see that the retriever returned irrelevant documents, or that the model ignored its system prompt, or that a tool call timed out and the agent hallucinated an answer instead. Traces also feed the eval loop. When you find a bad trace in production, you can turn it into a dataset record and add it to your eval suite, making sure the same failure is caught automatically in the future.