Yes, Phoenix tracing for autogen-agents is currently verbose by design: it captures all prompt/response pairs, agent interactions, and context passed between steps, to provide full visibility into multi-agent workflows. This includes detailed spans for each LLM call and agent action, which can result in a large number of trace entries, especially in prompt chaining or group chat scenarios. This verbosity is intentional to aid debugging and performance analysis, but some users have noted it can be overwhelming and there are ongoing discussions about reducing redundant messages and improving clarity in future releases. See <https://github.com/Arize-ai/openinference/issues/1687|autogen instrumentation followup> and <https://arize.com/docs/phoenix/integrations/frameworks/autogen/autogen-tracing#install|AutoGen tracing docs> for more context.
If you want to reduce verbosity, currently there is no built-in flag to limit trace detail for autogen-agents in Phoenix, but you can filter or post-process traces in your observability dashboard. Would you like more detail on how to interpret or filter these traces?
Sources:
- <https://github.com/Arize-ai/openinference/issues/1687|autogen instrumentation followup>
- <https://arize.com/docs/phoenix/integrations/frameworks/autogen/autogen-tracing#install|AutoGen tracing docs>
