AI-901 ยท Building agents in Foundry
How do you test and refine an agent?
Using the agents playground to try conversations, inspect tool calls and traces, refine instructions and spot wrong tool use or ungrounded answers.
Key points
- Chat with it in the agents playground, then open the trace to see inputs, outputs, tool calls and results, token use and latency for each step.
- If it answers without using a tool it should have used, uses the wrong tool, or states things the sources do not support, adjust the instructions or tools and retest.
Source: https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept
Practice questions
Where in the Foundry portal can you chat with an agent to try it out before building a client app?
- The agents playground
- The billing page
- The model catalog filters
- The project networking settings
Show answer
Answer: A. The agents playground
The agents playground is the interactive place to chat with an agent and test its behavior and tools.
Which details can an agent trace in Foundry show you?
- Which tools were called and what they returned
- Token consumption of the run
- The documents the model was originally trained on
- Duration and latency of the steps
- A refund of the run's cost
Show answer
Answer: A. Which tools were called and what they returned, B. Token consumption of the run, D. Duration and latency of the steps
Traces capture inputs, outputs, tool calls and results, token consumption and timing. They do not reveal the model's training data or change billing plans.
True or false: if an agent's answer is fluent and confident, you can assume it is grounded in your data.
Show answer
Answer: False
False. Fluent text can still be invented. Check the trace for a tool call and compare the answer with the retrieved content.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
