Skip to main content

​
Spans

A span is the smallest unit of captured work in Latitude. Each span represents one operation inside an agent run: an LLM call, a tool call, a retrieval step, an HTTP request, or any custom operation you instrument. One or more spans make up a trace.

​
What a span captures

A span can record:
  • input and output content, when available
  • start time, end time, and duration
  • success or error status
  • tags and metadata
  • provider, model, token usage, and cost for LLM calls

​
Common span kinds

  • LLM call: a request to a model, with its prompt, response, token usage, and cost.
  • Tool call: a function or tool the agent invoked, with its input arguments and result. See Tool calls.
  • Retrieval: a lookup against a vector store, database, or search index.
  • HTTP request: an outbound call to an external service.
  • Custom: any operation you instrument yourself.

​
How spans form a trace

Spans nest to show the execution path of a single interaction, and a trace groups them into one complete interaction: 
Trace: user turn
├─ Span: retrieve context
├─ Span: LLM call
└─ Span: tool call
Open any trace to see its spans as a tree, then expand a span to inspect its input, output, timing, and status. Use spans when you need to understand exactly what happened inside a trace, such as where latency, cost, or an error came from.

​
Next steps

  • Traces: the complete interaction a span belongs to
  • Tool calls: inspect the tools the agent called
  • Sessions: group related traces into a conversation
  • Core Concepts: the full product vocabulary