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 or Arize AX.

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 or Arize AX

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

Attribute Mapping

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

More Info

• More info on OpenInference and Phoenix
• More info on OpenInference and Arize AX

License

Apache-2.0