Maf Tracing

Enable Tracing & Logging for MAF Workflows

Configure OpenTelemetry-based tracing for Microsoft Agent Framework workflows, exporting to Azure Application Insights and/or a generic OTLP endpoint.

Triggers

Activate this skill when the user wants to:

• Enable tracing or logging for a MAF workflow
• Export telemetry to Azure Application Insights
• Export traces to an OTLP-compatible endpoint (Jaeger, Zipkin, Grafana Tempo, etc.)
• Add observability or monitoring to an agent-framework project
• Debug or inspect executor/agent/LLM spans in a MAF workflow

Background

MAF automatically emits OpenTelemetry spans for every executor invocation, agent call, and LLM request. No instrumentation changes are needed inside Executor classes. You only need to:

1. Configure an exporter — where spans are sent (Application Insights, OTLP endpoint, or both)
2. Call configure_otel_providers() — activates MAF's built-in instrumentation

This must happen once at application startup, before any workflow.run() calls.

Rules

1. Read the user's project first — Check for an existing requirements.txt, .env, and entry-point script (e.g., main.py, app.py, run_*.py).
2. Ask what export destination(s) are needed — Application Insights, OTLP endpoint, or both. If unclear, default to both.
3. Do not modify Executor classes — MAF tracing is automatic. Never add manual tracer.start_span() calls inside @handler methods unless the user explicitly asks for custom spans.
4. Call order matters — Exporter configuration must happen BEFORE configure_otel_providers(), and both must happen BEFORE any workflow.run().
5. Add packages to requirements.txt — Append only the packages the user needs (see Packages section). Do not duplicate existing entries.
6. Use environment variables — Never hardcode connection strings or endpoints. Always read from os.environ or .env.
7. Generate .env.example — Provide a template showing which environment variables are required.
8. Python logging integration — When the user asks for logging (not just tracing), also configure the Python logging module to export via OpenTelemetry using opentelemetry-sdk log handler.

Environment Variables

OTEL_EXPORTER_OTLP_TRACES_ENDPOINT takes precedence over OTEL_EXPORTER_OTLP_ENDPOINT for traces.

Packages

Setup Patterns

Pattern A: Application Insights Only

Use when the user wants to send traces to Azure Application Insights.

Required env var: APPLICATIONINSIGHTS_CONNECTION_STRING

Required packages:

azure-monitor-opentelemetry>=1.6.4

Setup code (add at the top of the entry-point script, before workflow.run()):

import os
from dotenv import load_dotenv
from azure.monitor.opentelemetry import configure_azure_monitor
from agent_framework.observability import configure_otel_providers

load_dotenv()

# Step 1: Configure Azure Monitor exporter (traces, metrics, logs → App Insights)
configure_azure_monitor(
    connection_string=os.environ["APPLICATIONINSIGHTS_CONNECTION_STRING"]
)

# Step 2: Enable MAF's built-in instrumentation (executor, agent, LLM spans)
configure_otel_providers()

Pattern B: OTLP Endpoint Only

Use when the user wants to send traces to a generic OTLP-compatible backend (Jaeger, Grafana Tempo, Aspire Dashboard, etc.).

Required env var: OTEL_EXPORTER_OTLP_ENDPOINT

Required packages:

opentelemetry-sdk>=1.25.0
opentelemetry-exporter-otlp-proto-http>=1.25.0

Setup code:

import os
from dotenv import load_dotenv
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.sdk.resources import Resource
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from agent_framework.observability import configure_otel_providers

load_dotenv()

# Step 1: Set up the OTLP exporter with a TracerProvider
resource = Resource.create({
    "service.name": os.environ.get("OTEL_SERVICE_NAME", "maf-workflow"),
})
tracer_provider = TracerProvider(resource=resource)

otlp_exporter = OTLPSpanExporter(
    endpoint=os.environ.get("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT")
    or os.environ.get("OTEL_EXPORTER_OTLP_ENDPOINT"),
)
tracer_provider.add_span_processor(BatchSpanProcessor(otlp_exporter))
trace.set_tracer_provider(tracer_provider)

# Step 2: Enable MAF's built-in instrumentation
configure_otel_providers()

Pattern C: Both Application Insights and OTLP

Use when the user wants dual export — Application Insights for Azure-native monitoring plus an OTLP backend for local/third-party observability.

Required env vars: APPLICATIONINSIGHTS_CONNECTION_STRING, OTEL_EXPORTER_OTLP_ENDPOINT

Required packages:

azure-monitor-opentelemetry>=1.6.4
opentelemetry-sdk>=1.25.0
opentelemetry-exporter-otlp-proto-http>=1.25.0

Setup code:

import os
from dotenv import load_dotenv
from azure.monitor.opentelemetry import configure_azure_monitor
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from agent_framework.observability import configure_otel_providers

load_dotenv()

# Step 1a: Configure Azure Monitor (sets up its own TracerProvider internally)
configure_azure_monitor(
    connection_string=os.environ["APPLICATIONINSIGHTS_CONNECTION_STRING"]
)

# Step 1b: Add OTLP exporter to the existing TracerProvider
otlp_endpoint = (
    os.environ.get("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT")
    or os.environ.get("OTEL_EXPORTER_OTLP_ENDPOINT")
)
if otlp_endpoint:
    otlp_exporter = OTLPSpanExporter(endpoint=otlp_endpoint)
    tracer_provider: TracerProvider = trace.get_tracer_provider()
    tracer_provider.add_span_processor(BatchSpanProcessor(otlp_exporter))

# Step 2: Enable MAF's built-in instrumentation
configure_otel_providers()

Pattern D: Python Logging via OpenTelemetry

Use when the user also wants Python logging calls to be exported alongside traces.

Additional packages (on top of Pattern A, B, or C):

opentelemetry-sdk>=1.25.0

Setup code (add after the exporter setup, before configure_otel_providers()):

import logging
from opentelemetry.sdk._logs import LoggerProvider
from opentelemetry.sdk._logs.export import BatchLogRecordProcessor
from opentelemetry._logs import set_logger_provider

# If using Application Insights, configure_azure_monitor() already handles log export.
# If using OTLP only, set up the OTLP log exporter:
from opentelemetry.exporter.otlp.proto.http._log_exporter import OTLPLogExporter

logger_provider = LoggerProvider(resource=resource)
logger_provider.add_log_record_processor(
    BatchLogRecordProcessor(OTLPLogExporter(
        endpoint=os.environ.get("OTEL_EXPORTER_OTLP_ENDPOINT"),
    ))
)
set_logger_provider(logger_provider)

# Bridge Python logging to OpenTelemetry
from opentelemetry.instrumentation.logging import LoggingInstrumentor
LoggingInstrumentor().instrument(set_logging_format=True)

Note: When using azure-monitor-opentelemetry (Pattern A/C), configure_azure_monitor() already captures Python logs by default. The above is only needed for OTLP-only setups.

Implementation Steps

Step 1: Determine export destination(s)

Ask the user or infer from context:

• Application Insights → Pattern A
• OTLP endpoint → Pattern B
• Both → Pattern C

If the user says "tracing" without specifying a destination, default to Pattern C (both).

Step 2: Update requirements.txt

Append the required packages. Do not duplicate existing entries. Example additions for Pattern C:

azure-monitor-opentelemetry>=1.6.4
opentelemetry-sdk>=1.25.0
opentelemetry-exporter-otlp-proto-http>=1.25.0

Step 3: Create or update .env.example

Add the environment variables relevant to the chosen pattern:

# === Tracing & Observability ===
# Application Insights (Azure Portal → App Insights → Overview → Connection String)
APPLICATIONINSIGHTS_CONNECTION_STRING=InstrumentationKey=xxx;IngestionEndpoint=https://xxx.in.applicationinsights.azure.com/

# OTLP endpoint (e.g., Jaeger, Grafana Tempo, Aspire Dashboard)
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318

# Optional: override service name in trace backends
OTEL_SERVICE_NAME=my-maf-workflow

Step 4: Add tracing setup to entry-point script

Insert the setup code from the appropriate pattern at the top of the entry-point script (after imports, before any workflow.run() call). The setup must execute once at module load / application startup.

Placement rules:

• If the entry point uses asyncio.run(main()), place setup code inside main() before workflow.run().
• If the entry point is a module-level script, place setup code after imports and load_dotenv().
• If the entry point is a web server (FastAPI, Flask), place setup code in the application factory or startup event.

Step 5: Verify

1. Run the workflow and check that traces appear in the configured destination.
2. For Application Insights: Azure Portal → Application Insights → Transaction Search → filter by "Dependency" or "Request".
3. For OTLP: Check the collector backend UI (e.g., Jaeger UI at http://localhost:16686).

Gotchas

1. Call order matters — configure_azure_monitor() and/or OTLP exporter setup must happen BEFORE configure_otel_providers(). If reversed, MAF spans won't be exported.
2. configure_otel_providers() must run BEFORE workflow.run() — Otherwise, executor-level spans are not generated.
3. Do not call setup code inside Executors — Tracing setup is application-level, not per-request. Calling it inside a @handler method will create duplicate exporters and corrupt traces.
4. configure_azure_monitor() creates its own TracerProvider — When combining with OTLP (Pattern C), add the OTLP exporter to the existing provider via trace.get_tracer_provider() rather than creating a new TracerProvider.
5. Missing configure_otel_providers() — Without this call, you'll see Application Insights or OTLP infrastructure telemetry but no MAF-specific spans (executor transitions, agent calls, LLM requests).
6. Connection string format — The APPLICATIONINSIGHTS_CONNECTION_STRING starts with InstrumentationKey= followed by a GUID. Do not confuse it with the Instrumentation Key alone.
7. OTLP endpoint trailing path — OTEL_EXPORTER_OTLP_ENDPOINT should be the base URL (e.g., http://localhost:4318). The SDK appends /v1/traces automatically. Do not include /v1/traces in the env var.
8. gRPC vs HTTP — Default protocol is http/protobuf (port 4318). If the collector uses gRPC (port 4317), set OTEL_EXPORTER_OTLP_PROTOCOL=grpc and install opentelemetry-exporter-otlp-proto-grpc instead.

Example: Complete Entry Point with Tracing

"""Entry point for a MAF workflow with full tracing setup."""
import asyncio
import os

from dotenv import load_dotenv

load_dotenv()


def setup_tracing():
    """Configure telemetry exporters and MAF instrumentation. Call once at startup."""
    from agent_framework.observability import configure_otel_providers

    # Application Insights
    appinsights_conn = os.environ.get("APPLICATIONINSIGHTS_CONNECTION_STRING")
    if appinsights_conn:
        from azure.monitor.opentelemetry import configure_azure_monitor
        configure_azure_monitor(connection_string=appinsights_conn)

    # OTLP endpoint (optional, additive)
    otlp_endpoint = (
        os.environ.get("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT")
        or os.environ.get("OTEL_EXPORTER_OTLP_ENDPOINT")
    )
    if otlp_endpoint:
        from opentelemetry import trace
        from opentelemetry.sdk.trace import TracerProvider
        from opentelemetry.sdk.trace.export import BatchSpanProcessor
        from opentelemetry.sdk.resources import Resource
        from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter

        otlp_exporter = OTLPSpanExporter(endpoint=otlp_endpoint)
        tracer_provider = trace.get_tracer_provider()

        # If Azure Monitor already set a TracerProvider, reuse it; otherwise create one
        if not isinstance(tracer_provider, TracerProvider):
            resource = Resource.create({
                "service.name": os.environ.get("OTEL_SERVICE_NAME", "maf-workflow"),
            })
            tracer_provider = TracerProvider(resource=resource)
            trace.set_tracer_provider(tracer_provider)

        tracer_provider.add_span_processor(BatchSpanProcessor(otlp_exporter))

    # Enable MAF's built-in spans (must be last)
    configure_otel_providers()


async def main():
    setup_tracing()

    # Import and run your workflow here
    from workflow import create_workflow

    workflow = create_workflow()
    result = await workflow.run("Hello, world!")
    print(result.get_outputs())


if __name__ == "__main__":
    asyncio.run(main())