AI-103 Practice Question 5: Agent Tracing Latency
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 1 · question 5 of 10
A prompt agent in production has four tools attached, and one of those tools calls a second internal service of its own. Over the past week its 95th-percentile response time has climbed from three seconds to twenty-two, but only on some conversations; most still return promptly. The team can reproduce nothing locally and needs to see, for one specific slow conversation, the order in which the model calls and tool invocations actually ran and how long each one took, including the nested call.
What should they set up?
Show answer and explanation
Answer:
Tracing is the Foundry observability feature built for this question; Microsoft frames it as answering "Which step introduced an error or latency spike?". A trace captures an agent run's user inputs and outputs, tool usage including tool calls and results, token consumption, and time signals such as duration and latency. Its building blocks are spans, each recording start and end times and attributes, and spans nest, so you see the full call stack and sequence of operations, exactly what is needed when a tool invokes another service. Traces are exported to the Application Insights resource connected to the project, and tracing is generally available for prompt and hosted agents.
- A is wrong because service metrics dashboards are aggregate. They confirm that latency rose across the population, which the team already knows, but cannot show which step inside one particular conversation stalled, and here most conversations are fine. - C is wrong because evaluations measure response *quality* against a dataset, scoring things like groundedness and relevance. They report nothing about per-step timing in a live production run, and a slow answer can be a perfectly good answer. - D is wrong because guardrails are safety and security controls. Their intervention points scan content for risks such as harmful material or prompt attacks and then block it; they are not a performance instrument, and for agents the only supported action is annotate and block.
Reference: Agent tracing overview - Microsoft Foundry