AI-300 Practice Question 28: Agent Tracing Spans
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 3 · question 8 of 10
A multi-step Foundry agent occasionally returns an answer built from out-of-date information and takes nine seconds instead of the usual two. The team can already see the final response and the total latency, but cannot tell which model call or tool invocation is responsible, and the agent's tools sometimes call other tools. What should you enable to attribute the delay and the stale content to a specific step?
Show answer and explanation
Answer:
Tracing "captures key details during an agent run, such as inputs, outputs, tool usage, retries, latencies, and costs", and it is built from spans: "Spans are the building blocks of traces, representing single operations within a trace. Each span captures start and end times, attributes, and can be nested to show hierarchical relationships, so you can see the full call stack and sequence of operations." Microsoft frames it against precisely this problem (an agent invoking a tool which invokes another tool), and says tracing helps answer "Where did this response come from?" and "Which step introduced an error or latency" (option B).
- A is wrong because continuous evaluation scores the quality and safety of responses on sampled production traffic. It can tell you answers are getting worse; it produces no per-step timing and cannot point at the tool call that took seven extra seconds. - C is wrong because a safety run on a curated dataset measures harm categories against test inputs. It never touches the slow production runs and says nothing about execution flow. - D is wrong because provisioned throughput addresses capacity contention on model inference. It cannot explain a stale answer at all, and without traces there is no evidence the model call is even the slow step.
Reference: Agent tracing overview - Microsoft Foundry