Skip to main content
In this lesson, you’ll learn to use a Phoenix trace to identify the information an agent received and check whether it supports the final answer. You’ll practice with Wonder Toys, a shopping assistant agent, by comparing its product search results with its recommendation.

​
Before you start

Continue with the request you sent in Add an agent span, asking for dinosaur toys for a 6-year-old under $25. To inspect how Wonder Toys handled it, open wonder-toys → Traces in Phoenix and select that request’s trace. In the screenshots, Wonder Toys recommends a toy that costs more than $25. Your results may be different, but you can follow the same steps with your own trace.

​
Step 1: Read the model’s search request

  • Select the first ChatAnthropic model call on the left.
  • Under Info on the right, read the shopper’s message in Input.
Phoenix trace with the first ChatAnthropic row selected on the left and the Info tab showing Input and the shopper's dinosaur toy request on the right

The selected model call shows the shopper's request under Input.

Then scroll to Output and find search-products. In this example, the model asks the tool for dinosaur toys for age 6.
First model call output showing search-products with query dinosaur toys and min_age and max_age both set to 6

The model's search request appears under Output.

  • If you don’t see search-products, check the next ChatAnthropic span.
  • If none of the model calls requested a search, skip to Step 4.

​
Step 2: Read the search results

  • Select the search-products row and check that Input matches the search from Step 1.
  • Then read Output to see the returned products.
In this example, the tool returned a Dinosaur Explorer Kit priced at $27.99. Fetching product records like these is called retrieval.
Phoenix search-products span showing the dinosaur query, age filters, and returned Dinosaur Explorer Kit priced at 27.99

The tool's Output shows the products returned by the search.

Next, check whether the model received these results.

​
Step 3: Compare the recommendation with the results

Select the next ChatAnthropic model call and find tool: search-products under Input. In this example, that message contains the kit’s $27.99 price.
Tool message in the final model call's input showing the Dinosaur Explorer Kit and its price of 27.99

The search results are included in the next model call's Input.

Scroll to Output and check the recommendation against the product details and the shopper’s requirements. Here, the model recommended the kit while noting that it was over budget.
Final model call output recommending the Dinosaur Explorer Kit at 27.99 and noting that it is over budget

The answer recommends the kit at $27.99 and notes that it is over budget.

  • If there are more searches, repeat this comparison until you reach the final answer.
  • If no products came back, check how the answer explains that.

​
Step 4: Explain the recommendation

For the example in the screenshots, you can now explain:
The model asked the tool for dinosaur toys for age 6 and received results that included a Dinosaur Explorer Kit priced at $27.99. It then recommended that kit while noting that it was over budget. The model had the price information but still included a toy above the shopper’s $25 limit.
Use your own trace to explain what happened, whether the answer was correct or not. If no search ran, compare the model’s input and answer.

​
Step 5: Check the cost of the request

You’ve seen how the agent reached its answer. Now let’s check how much its model calls cost.
  • Find the request total. Read Total Cost at the top of the trace, marked 1 below.
  • Compare the model calls. Select each model span, such as ChatAnthropic. Hover over the token count or cost beneath its name, marked 2, to see the input and output breakdown. Tokens are pieces of content the model reads or generates.
  • Find the most expensive call. Read its Input and Output to see what information it received and what it produced.
Phoenix in light mode with the request's Total Cost outlined as 1 and a selected model call's tokens and cost outlined as 2. The open tooltip separates input and output tokens and costs.Two stacked crops from the same Phoenix trace: 1 shows the request's total cost; 2 shows the selected model call's token and cost breakdown, separated into input and output.

1: Total Cost covers the whole request. 2: Hover over a model call's token count or cost to inspect its breakdown.

Phoenix calculates model-call costs from recorded token usage and model pricing. Use Total Cost for the whole request.
Latency is the elapsed time shown at the top of the trace. Compare individual model and tool calls to find where the agent spent time. Use the trace latency for the whole request; parent and child durations overlap.

​
Continue the conversation

You’ve reviewed one request’s behavior and cost. Keep the Wonder Toys chat open, then follow a conversation with sessions to check how the agent uses an earlier answer in a follow-up.