Span
A single timed step inside a trace — one model call or tool invocation with its inputs and outputs.
Think of it like
One line item on an itemized receipt — this thing, this long, this cost.
Example
Within the trace, one span is the “search docs” tool call: it took 400ms, took this query, returned these three results.
How it actually works
Spans are the building blocks of a trace — each captures one unit of work with a start, end, and metadata, and spans nest to show what called what. They’re where you find the actual detail: the exact prompt sent, the tokens used, the tool’s raw response. Borrowed from distributed tracing, the span/trace pair is what lets you zoom from “the whole run” down to “this one slow call.”
For product teams
Spans are the zoom-in — where you see the specific step that cost the most time or money.
For engineers
A single timed, attributed unit of work within a trace, nestable to reflect call hierarchy.
Related
- Trace — The whole run a span belongs to.
- Agent Observability — What spans add up to enable.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome