Langfuse - Microsoft Agent Framework Issue

[eddieahn]

When sending data from Microsoft Agent Framework to Langfuse, should the invoke_agent span have gen_ai.request.model? My assumption is that its an orchestration span, not an LLM call so in Langfuse the issue below occurs:

Even though the Microsoft Agent Framework sets capture_usage=False, Langfuse will still attempt to infer token usage because:

• The span has gen_ai.request.model: "gpt-4.1", which maps to the model property
• This causes Langfuse to classify it as a generation type observation
• When usage details aren't provided and a model is specified, Langfuse uses its predefined tokenizers to estimate tokens
• The Agent Framework's invoke_agent span gets token counts inferred by Langfuse
• Any nested actual LLM calls (like the OpenAI completion) also get their own token counts
• Langfuse sums these together in the total

Is this an issue that needs to be fixed on the Langfuse side or can we remove that specific field from being passed?

[eavanvalkenburg]

Thanks for reaching out @eddieahn i did a quick check in the otel docs and it specifically has request model also under the invoke agent span: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-agent-spans/#invoke-agent-span so in this case, I think it's up to langfuse to ensure they follow the setup and not infer token usage when you configure that,when I tested with local langfuse it did seem to work correctly though...

[eddieahn]

@eavanvalkenburg Hmm I see, thanks for the quick reply. This is what I see on my end. You can see that both the invoke and gpt-4.1 has tokens associated with it and total token and cost represents all of those summed up. Is that different than what you see on your end?