openinference-instrumentation-agent-framework 0.1.5

OpenInference span processor for Microsoft Agent Framework - transforms native OpenTelemetry spans to OpenInference format

OpenInference Microsoft Agent Framework Instrumentation

OpenInference span processor for Microsoft Agent Framework that transforms native OpenTelemetry spans to OpenInference format for compatibility with OpenInference-compliant backends like Arize Phoenix.

Tested with Agent Framework core/OpenAI packages 1.0.0

Installation

pip install openinference-instrumentation-agent-framework

Install with Microsoft Agent Framework core and OpenAI support:

pip install openinference-instrumentation-agent-framework[instruments]

Overview

Microsoft Agent Framework emits telemetry using GenAI semantic conventions (gen_ai.* attributes). This package provides a SpanProcessor that transforms these spans to OpenInference format, enabling compatibility with observability tools that support the OpenInference standard.

Note: This instrumentation targets Agent Framework's stable 1.0.0+ API surface and tracks later stable releases in CI.

Usage

Basic Setup

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, SimpleSpanProcessor
from agent_framework.observability import enable_instrumentation
from openinference.instrumentation.agent_framework import (
    AgentFrameworkToOpenInferenceProcessor
)

# Create tracer provider
tracer_provider = TracerProvider()

# Add OpenInference processor to transform spans
tracer_provider.add_span_processor(
    AgentFrameworkToOpenInferenceProcessor()
)

# Add exporter (Console for demo, use OTLP for Phoenix)
tracer_provider.add_span_processor(
    SimpleSpanProcessor(ConsoleSpanExporter())
)

# Set as global tracer provider
trace.set_tracer_provider(tracer_provider)

# Enable agent-framework instrumentation
enable_instrumentation(enable_sensitive_data=True)

# Use framework normally - spans will be transformed automatically
from agent_framework.openai import OpenAIChatClient

client = OpenAIChatClient(model="gpt-4o-mini", api_key="your-key")
agent = client.as_agent(name="Assistant", instructions="You are helpful.")
response = await agent.run("Hello!")

With Phoenix (Arize)

from opentelemetry import trace
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from agent_framework.observability import enable_instrumentation
from openinference.instrumentation.agent_framework import (
    AgentFrameworkToOpenInferenceProcessor
)

# Configure Phoenix endpoint
endpoint = "http://localhost:6006/v1/traces"

# Setup tracer with OpenInference processor
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(AgentFrameworkToOpenInferenceProcessor())
tracer_provider.add_span_processor(
    SimpleSpanProcessor(OTLPSpanExporter(endpoint=endpoint))
)
trace.set_tracer_provider(tracer_provider)

# Enable instrumentation
enable_instrumentation(enable_sensitive_data=True)

# Your agent code here

Transformation Details

Span Kind Mapping

MS Agent Framework Operation	OpenInference Span Kind
chat	LLM
execute_tool	TOOL
invoke_agent	AGENT
workflow.run	CHAIN
executor.process	CHAIN

Attribute Mapping

Source (GenAI)	Target (OpenInference)
gen_ai.request.model	llm.model_name
gen_ai.provider.name	llm.provider
gen_ai.usage.input_tokens	llm.token_count.prompt
gen_ai.usage.output_tokens	llm.token_count.completion
gen_ai.input.messages	llm.input_messages.* (flattened)
gen_ai.output.messages	llm.output_messages.* (flattened)
gen_ai.tool.name	tool.name
gen_ai.tool.call.id	tool.call_id
gen_ai.tool.call.arguments	tool.parameters
gen_ai.conversation.id	session.id

Message Format Transformation

MS Agent Framework messages:

{
  "role": "user",
  "parts": [
    {"type": "text", "content": "Hello"}
  ]
}

Transformed to OpenInference flattened format:

llm.input_messages.0.message.role = "user"
llm.input_messages.0.message.content = "Hello"

Debug Mode

Enable debug mode to log transformation details:

processor = AgentFrameworkToOpenInferenceProcessor(debug=True)

Requirements

• Python >= 3.10, < 3.15
• opentelemetry-api >= 1.39.0
• opentelemetry-sdk >= 1.39.0
• openinference-semantic-conventions >= 0.1.25
• agent-framework-core >= 1.0.0 (optional, install with [instruments] extra)
• agent-framework-openai >= 1.0.0 (optional, install with [instruments] extra)

Important Notes

Agent Framework API Stability

Microsoft Agent Framework's stable OpenAI client changed meaningfully at 1.0.0:

• This instrumentation is tested against agent-framework-core==1.0.0 and agent-framework-openai==1.0.0
• The -latest test variant tracks breaking changes in new releases
• If you encounter API compatibility issues, pin to the tested stable version:

  pip install agent-framework-core==1.0.0 agent-framework-openai==1.0.0
  

If you have existing VCR cassettes from the 1.0.0rc1 line, re-record them after upgrading. The stable OpenAIChatClient uses the OpenAI Responses API rather than the older Chat Completions path.

Sensitive Data

Set enable_sensitive_data=True when calling enable_instrumentation() to capture message content in traces. This is required for full observability but may include PII.

Development

Running Tests

Tests use VCR cassettes to replay recorded API interactions:

# Run with recorded cassettes (no API key needed)
pytest tests/test_processor.py -v

# Re-record cassettes (requires OPENAI_API_KEY)
export OPENAI_API_KEY=your_key
rm -rf tests/cassettes/
pytest tests/test_processor.py -v --record-mode=rewrite

License

Apache-2.0