Simple Calculator with Observability and Tracing

Why it matters

Root-cause diagnosis depends on knowing which step invoked which tool and model. This example shows how the Agent Toolkit records that structure, and how the exporter is a configuration choice.

What to practise

Start a local Phoenix server and run the workflow with nat run and config-phoenix.yml, whose general.telemetry.tracing block names the exporter. Then run config-phoenix-nested.yml: the power_of_two tool calls calculator__multiply, and filtering on nat.function.parent_name separates the tools the react_agent chose from nested calls. Export the same run to a JSON Lines file with config-otel-file.yml. Optionally, switch to another backend config, such as MLflow or Langfuse, and compare how each backend presents the same run.

How it fits

The parent-child hierarchy is the evidence 8.2.3 reads to isolate a cause. Joining those spans to service logs and inference requests is 8.2.1’s problem, which a single tracing backend does not solve for you.