Hi team, I’m running Arize Phoenix with the autogen agent framework, and I’ve noticed that each test run shows some unexpected “unknown” spans in addition to the normal chain spans. I’m not sure if this is due to the code structure or configuration. Do you have any recommendations on how to investigate or clean them up? thank you!
"Unknown" spans in Phoenix with the Autogen agent framework typically result from the auto-instrumentation wrapping many internal methods, which can create extra or less-informative spans. There is currently no built-in configuration to filter or group these spans at the instrumentation level; each wrapped method call results in a separate span, and "unknown" kind appears when the span type can't be determined from context or attributes. To reduce or clean up these spans, you can use the suppress_tracing context manager to manually suppress tracing in specific code blocks, but this requires explicit intervention in your code. For deeper investigation, review which methods are being wrapped and consider if you can limit instrumentation scope or add more semantic attributes to clarify span types. Otherwise, changes to span grouping or filtering would require upstream changes to the instrumentation logic itself.
Sources:
- Autogen Agentchat tracing verbosity and span kinds
💡 Hint: Mention RunLLM in the thread for followups.
Hi anki y.! Thanks for bringing this up. Just tried this out with the RoundRobinGroupChat and got the same unknown spans. Seems like an issue with our instrumentor. In the meantime, I would suggest suppress_tracing (https://arize.com/docs/phoenix/tracing/how-to-tracing/advanced/suppress-tracing). I suppressed tracing for the lines creating the AssistantAgent and RoundRobinGroupChat and that did the trick. My code snippet: from openinference.instrumentation import suppress_tracing with suppress_tracing(): primary_agent = AssistantAgent( "primary", model_client=model_client, system_message="You are a helpful AI assistant.", ) text_termination = TextMentionTermination("APPROVE") team = RoundRobinGroupChat( [primary_agent, critic_agent], termination_condition=text_termination )
If you're still getting unknown spans, let me know!
Thanks for reporting anki y.. Sanjana Y. can you file us an issue on GitHub with the repro?
