What is Agent tracing?

There is no single formal specification for agent tracing. The cited page is the OpenTelemetry description of traces in general. It does not mention agents, so applying it to an agent run is an interpretation: each model call or tool call becomes one unit of work that can be recorded as a span.

According to the page, a trace shows the full path of a request and is made of spans that share a trace ID. A span represents one unit of work and carries a name, a parent ID that is empty for the root span, start and end times, attributes, events, links and a status. Attributes are key-value metadata describing the operation. Events are timestamped notes on a moment inside the span. Links connect a span to causally related spans, including ones in a different trace, which helps with asynchronous work.

Spans in different processes or machines can still join one trace because the span context, which holds the trace ID and span ID, is passed between services. Parent and child relationships give the trace its hierarchy, so a tool call can be seen as a child of the step that triggered it. Span status is Unset by default, Error when something failed, or Ok when a developer marks success explicitly.