A span is a single unit of work inside a trace - one model call, one retrieval step, or one tool invocation. Spans carry their own timing, cost, and attributes, so a trace becomes a waterfall you can read to find which step is slow or expensive.
Why it matters
Knowing a trace was slow is less useful than knowing which step was slow. Spans localize time and cost inside a workflow.
How it works
Each model call, retrieval step, or tool invocation is a span with its own timing, cost, and attributes. Stacked in order, spans form the waterfall that makes a trace readable.