Tracing and Trajectory Logging

Every agent run can leave behind a structured record that holds each model call, each tool call, each token count, and each duration, with every entry linked to the step that caused it (a trace). Each timed entry inside that record is a span. A model call is one span, and a tool call is another.

An ordinary log can tell you that a request failed. An agent's code path was chosen by a model, so the reasons for a result lie in what the model received and asked for. What must a trace record so that someone who did not watch a run can explain it?