[bug] Agno OpenInference instrumentation: output.value missing in streaming mode ⇒ "undefined" in Langfuse

[rattanjot-statisfy]

Describe the bug

When using openinference-instrumentation-agno with Agno agents in streaming mode, the OpenInference instrumentation sometimes fails to set the output.value span attribute on Agent.run spans.

Because output.value is missing, Langfuse (via OTLP) displays the output as undefined, even though the agent produces a valid final answer.

Root cause appears to be that the instrumentation only uses RunOutput.content and only sets output.value when content is non‑None. In streaming mode content can be None or incomplete while the real output is available via other fields on RunOutput (e.g. get_content_as_string() / messages).

---

To Reproduce

1. Install:

   agno==2.3.2
   openinference-instrumentation-agno==0.1.23
   opentelemetry-sdk
   opentelemetry-exporter-otlp-proto-http
   2. Configure OTLP → Langfuse (or any OTLP backend):

   from opentelemetry import trace
   from opentelemetry.sdk.trace import TracerProvider
   from opentelemetry.sdk.trace.export import SimpleSpanProcessor
   from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
   from openinference.instrumentation.agno import AgnoInstrumentor

   tracer_provider = TracerProvider()
   tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter()))
   trace.set_tracer_provider(tracer_provider)

   AgnoInstrumentor().instrument()
   3. Run an Agno agent with streaming:

   from agno.agent import Agent
   from agno.models.google import Gemini
   from agno.tools import tool
   import os

   @tool
   def add(a: float, b: float) -> float:
   return a + b

   model = Gemini(
   id="gemini-2.5-flash",
   project_id=os.getenv("GOOGLE_CLOUD_PROJECT_ID"),
   location=os.getenv("GOOGLE_CLOUD_LOCATION", "us-central1"),
   vertexai=True,
   include_thoughts=True,
   )

   agent = Agent(
   model=model,
   tools=[add],
   reasoning=True,
   markdown=True,
   debug_mode=True,
   )

   for chunk in agent.run(
   "I need to calculate: Start with 100, add 25, then subtract 15",
   stream=True,
   ):
   pass
   4. Look at the exported spans in:

   • Langfuse (OTLP traces), or
   • A custom OTLP/console exporter.

2. Inspect the Agent.run span:

   • output.mime_type is set
   • But output.value is often missing or empty for some runs.

---

Expected behavior

• output.value (SpanAttributes.OUTPUT_VALUE, semantic name output.value) should always be set on Agent.run spans to a string representation of the final RunOutput (or empty string at minimum).
• Downstream backends like Langfuse should never see “undefined” when a final answer exists; they should see the model/agent output.

This matches OpenInference’s semantic conventions for LLM / agent traces.

---

Screenshots

• Langfuse traces show:
  • Inputs correctly populated
  • Spans for Agent.run present
  • Output field displayed as undefined because output.value is not present on the span.

(You can attach your actual screenshots here.)

---

Desktop (please complete the following information):

• OS: macOS (e.g. Sonoma 14.x)
• Python: 3.11.x
• Agno: 2.3.2
• openinference-instrumentation-agno: 0.1.23
• opentelemetry-sdk: (version you’re using)
• opentelemetry-exporter-otlp-proto-http: (version you’re using)

---

Additional context

Current implementation (problematic)
In _runs_wrapper.py:

def _extract_run_response_output(run_response: Union[RunOutput, TeamRunOutput]) -> str:
if run_response and run_response.content:
if isinstance(run_response.content, str):
return run_response.content
else:
return str(run_response.content.model_dump_json())
return ""And in run_stream:

if run_response is not None:
if run_response.content is not None:
span.set_attribute(OUTPUT_VALUE, _extract_run_response_output(run_response))
span.set_attribute(OUTPUT_MIME_TYPE, JSON)
span.set_status(trace_api.StatusCode.OK)Issues:

• RunOutput.content can be None in streaming mode even when there is a valid final output.
• RunOutput exposes other ways to get the output:
  • get_content_as_string()
  • messages (assistant messages)
• Because of the if run_response.content is not None guard, output.value is sometimes never set.

Observed behavior with a debug exporter

Using a custom SpanExporter, I can see:

• Some Agent.run spans have:

  output.value: ""
  output.mime_type: "application/json"

  • Other Agent.run spans have:

  output.value: "" # or completely missing
  output.mime_type: "application/json"
  Those empty/missing cases correspond to the “undefined” output in Langfuse.

Proposed direction

• Update _extract_run_response_output to:

  • Try content
  • Fallback to get_content_as_string()
  • Fallback to messages (last assistant message)
  • Always return a string (possibly empty).

• Update run_stream / arun_stream to:

  • Always set OUTPUT_VALUE (no if run_response.content is not None guard).
  • Optionally fall back to instance.get_last_run_output() when run_response is None or has empty content.

I have a working local change that fixes the issue end‑to‑end (Langfuse no longer shows “undefined”) and I’m happy to open a PR if this direction makes sense.

[dosubot]

Hi @rattanjot-statisfy! I'm Dosu and I’m helping the openinference team.

This bug is already tracked in a relevant open issue (#2451), and your analysis matches the root cause: in streaming mode, the instrumentation sometimes fails to set output.value because it doesn't always yield or capture the final RunOutput object, especially if yield_run_output isn't set correctly. This leads to blank or missing output in downstream systems like Langfuse [source].

There are open PRs addressing this, notably #2421, which updates the streaming wrappers to properly handle yield_run_output and yield_run_response, ensuring the final output is always available for instrumentation and that output.value is set as expected. The fix aligns with your proposed direction: always set output.value using the best available output representation, and avoid gating on content alone [source].

If you want an immediate workaround, explicitly pass yield_run_output=True when running your agent in streaming mode. This ensures the final output is yielded and traced [source].

Your local change and suggestion to improve output extraction logic are consistent with the direction of the open PRs. If you have further improvements—like more robust fallback logic for output extraction—they would be welcome as a PR, especially if they help cover edge cases not addressed by the current fixes.

To track resolution, monitor the status of PRs #2421 and #2450. Once merged, updating to the latest release should resolve the undefined output issue in Langfuse and other OTLP backends.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}