Trace Hierarchy
Run → Step → Trace. A run contains multiple steps. Each step has one trace with the complete execution record. The trace includes three timing markers —
queuedAt (when the step became ready), startedAt (when the agent began work), and completedAt (when results were returned).Conversation Logs
The full agent conversation for each step, including:- System prompts — the instructions provided to the agent
- User messages — the step input derived from the workflow definition and prior step outputs
- Assistant responses — the agent’s reasoning and generated output
- Tool use — tool calls made during execution with arguments and return values
Debugging with Traces
Traces answer the three core debugging questions:| Question | Where to Look |
|---|---|
| ”Why did this step fail?” | The error field on the step result, plus the last assistant message in the conversation log |
| ”Why was this step slow?” | The timing breakdown — compare queuedAt → startedAt (queue wait) vs startedAt → completedAt (execution time) |
| “What did the agent decide?” | The full conversation log shows the agent’s reasoning chain, tool calls, and output construction |
Each step trace also carries token attribution — input tokens, output tokens, cached tokens, model used, and cost at capture time. This links the trace directly to the analytics pipeline.
In Burgundy: Inspect traces from the observe page. →