To ensure each execution of your LangGraph workflow appears as a single trace (with all LLM/tool calls as child spans) in the Phoenix dashboard, you must invoke the workflow in a way that maintains a single root span per run. With Phoenix's OpenInference auto-instrumentation, this is achieved by calling your workflow as a callable (e.g., `workflow(args)`) rather than using `.forward(args)`, which can flatten or separate traces. This ensures all node executions and LLM calls are grouped as spans under one trace for each workflow run, making the trace tree clear in the UI. No manual instrumentation is required if you follow this pattern and have auto-instrumentation enabled for LangGraph/LangChain.
If you still see jumbled traces, double-check that you are not breaking context propagation (e.g., by using non-callable invocation patterns or by running nodes in separate processes without context propagation). For more, see the <https://arize.com/docs/phoenix/integrations/python/langgraph/langgraph-tracing|LangGraph Tracing guide> and this <https://arize-ai.slack.com/archives/C04R3GXC8HK/p1716308801.870979|Slack thread>.
Would you like more detail or a code example?
Sources:
- <https://arize.com/docs/phoenix/integrations/python/langgraph/langgraph-tracing|LangGraph Tracing - Phoenix Docs>
- <https://arize-ai.slack.com/archives/C04R3GXC8HK/p1716308801.870979|Slack Discussion: dspy/langgraph trace grouping>
