Verbose Tracing Output in Autogen-Agents with Phoenix: Is It Normal?

Zhuyi X.

·Jul 02, 2025 11:19 PM

I'm trying to use autogen-agents with phoenix, the tracing output looks really verbose, is that expected?

5 comments

· Sorted by Oldest

Zhuyi X.
·
also attached the script
RunLLM
APP
·
Yes, the current Phoenix tracing for AutoGen agents is intentionally verbose, capturing all agent interactions, message flows, and model calls for full observability. This includes detailed spans for each agent step, which can result in a large number of trace entries, especially in multi-agent or iterative workflows. This verbosity is a known design choice to help answer questions about agent actions, latency, token usage, and workflow structure, but there are ongoing discussions about reducing redundant or overly detailed spans in future updates (<https://github.com/Arize-ai/openinference/issues/1687|source>, <https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md|source>, <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md|source>). Would you like more detail on how to customize or filter the verbosity of traces? Sources: - <https://github.com/Arize-ai/openinference/issues/1687|https://github.com/Arize-ai/openinference/issues/1687> - <https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md|https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md> - <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md|https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md>
💡 Hint: Mention RunLLM in the thread for followups.
💡1
Zhuyi X.
·
By verbosity, I mean why are there so many traces, most of them with some IDs appended, the content feels hard to read. So you mean the verbosity isn't expected?
Can you check if your termination conditions are working as expected?
Yeah, it does seem working as expected. Here is a diff run of the same script, and the last assistant does say APPROVE at the end, e.g.
Zhuyi X.
·
btw, I'm not very familiar with autogen yet, but I learned that phoenix with autoinstrument can be quite a useful tool to understand how these agent frameworks works under the hood, so I'm trying out autogen + phoenix to get a sense of how autogen works.
Zhuyi X.
·
-

Zhuyi X.
·
also attached the script
RunLLM
APP
·
Yes, the current Phoenix tracing for AutoGen agents is intentionally verbose, capturing all agent interactions, message flows, and model calls for full observability. This includes detailed spans for each agent step, which can result in a large number of trace entries, especially in multi-agent or iterative workflows. This verbosity is a known design choice to help answer questions about agent actions, latency, token usage, and workflow structure, but there are ongoing discussions about reducing redundant or overly detailed spans in future updates (<https://github.com/Arize-ai/openinference/issues/1687|source>, <https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md|source>, <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md|source>). Would you like more detail on how to customize or filter the verbosity of traces? Sources: - <https://github.com/Arize-ai/openinference/issues/1687|https://github.com/Arize-ai/openinference/issues/1687> - <https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md|https://github.com/Arize-ai/phoenix/blob/main/docs/section-learn/agents/agent-workflow-patterns/autogen.md> - <https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md|https://github.com/Arize-ai/phoenix/blob/main/docs/tracing/integrations-tracing/autogen-support.md>
💡 Hint: Mention RunLLM in the thread for followups.
💡1
Zhuyi X.
·
By verbosity, I mean why are there so many traces, most of them with some IDs appended, the content feels hard to read. So you mean the verbosity isn't expected?
Can you check if your termination conditions are working as expected?
Yeah, it does seem working as expected. Here is a diff run of the same script, and the last assistant does say APPROVE at the end, e.g.
Zhuyi X.
·
btw, I'm not very familiar with autogen yet, but I learned that phoenix with autoinstrument can be quite a useful tool to understand how these agent frameworks works under the hood, so I'm trying out autogen + phoenix to get a sense of how autogen works.
Zhuyi X.
·
-

Verbose Tracing Output in Autogen-Agents with Phoenix: Is It Normal? | Arize AI Community