The customs-compliance agent occasionally returns a wrong HS code, and the team needs to see exactly which retrieval call, tool call, and intermediate LLM response led to the wrong output for one specific failing example. Which MLflow capability should they use to inspect that?
Select an answer to reveal the explanation.
Short Explanation
When you need to see inside one specific bad answer, step by step, you want tracing, not a rollback button and not a chart. MLflow tracing records each retrieval call, tool call, and LLM call as its own span within one agent run, so you can open exactly the failing example and watch the decisions unfold in order.
Full Explanation
MLflow tracing instruments an agent invocation and captures each internal step, such as a Vector Search retrieval call, a tool invocation like an HS-code lookup, and each LLM generation, as a span within a single trace tied to that one request. This is exactly what is needed to debug one specific wrong answer: the team can open the trace for that example and see, in order, what was retrieved, what the tool returned, and what the model generated at each stage, pinpointing where the wrong code was introduced. Model Registry aliases are a deployment mechanism for pointing a serving endpoint at a specific model version, such as rolling back to a previous version; they help recover from a regression but reveal nothing about why a specific example failed. Spark job metrics autologging captures cluster-level compute statistics like task duration and shuffle size, which is useful for performance tuning of distributed jobs but has no visibility into an agent's reasoning steps. Experiment comparison charts aggregate metrics across many separate runs to compare overall performance trends, which is the right tool for comparing prompt versions in bulk, as in the earlier scenario, but it is the wrong granularity for tracing one individual failing request end to end. Tracing and experiment tracking are complementary but distinct MLflow GenAI capabilities.