TRACING AND SPANS

WHAT NOT TO TRACE Not every detail deserves a span. Tiny internal calculations that take microseconds add clutter and cost without adding understanding. Trace the steps that a human would name when describing the run, such as asking the model, running a tool, or checking a result. Keep the picture simple enough that a new teammate can read it in a minute. TRACES AND EVALUATION TOGETHER Traces are not only for emergencies. You can also use them to improve quality over time. Collect traces of runs that users rated badly, study the common patterns, and add the patterns as new cases in your evaluation set. In this way the production system continually feeds the test suite with realistic examples. This loop, from trace to test to fix, is one of the most valuable habits in the whole field.