Reading the trace timeline

The trace timeline orders spans by their recorded sequence and displays their parent-child relationship.

A tool-calling trace may look like:

Agent
├── Model step 1
├── Tool: searchOrders
├── Model step 2
├── MCP tool: sendEmail
└── Model step 3

Select a span to inspect:

• Type, status, and name.
• Start time, end time, and duration.
• Provider and model.
• Parent span.
• Request and response payloads.
• Token categories and estimated cost.
• Metadata.
• Error type, message, and optional stack.

The waterfall is useful for streamed and tool operations with real start and end events. Non-stream model spans reconstructed at agent completion are marked in metadata and may have zero duration.

Failed and cancelled spans remain visible alongside successful operations so you can see where the workflow stopped.