LoongSuite CrewAI Instrumentation

This library provides automatic instrumentation for CrewAI, a framework for orchestrating role-playing, autonomous AI agents.

Installation

# Step 1: install LoongSuite distro
pip install loongsuite-distro

# Step 2 (Option C): install this instrumentation from PyPI
pip install loongsuite-instrumentation-crewai

Usage

from opentelemetry.instrumentation.crewai import CrewAIInstrumentor

# Instrument CrewAI
CrewAIInstrumentor().instrument()

# Your CrewAI code here
from crewai import Agent, Task, Crew

agent = Agent(
    role='Data Analyst',
    goal='Extract actionable insights',
    backstory='Expert in data analysis',
    verbose=True
)

task = Task(
    description='Analyze the latest AI trends',
    agent=agent
)

crew = Crew(
    agents=[agent],
    tasks=[task],
    verbose=True
)

result = crew.kickoff()

Telemetry

This instrumentation uses ExtendedTelemetryHandler from opentelemetry-util-genai and emits:

• enter_ai_application_system spans for Crew.kickoff and Flow kickoff entry points (gen_ai.span.kind=ENTRY, gen_ai.operation.name=enter)
• invoke_agent spans for CrewAI task and agent execution (gen_ai.span.kind=AGENT, gen_ai.operation.name=invoke_agent)
• execute_tool spans for CrewAI tool execution (gen_ai.span.kind=TOOL, gen_ai.operation.name=execute_tool)

CrewAI-specific framework details are kept in gen_ai.crewai.* attributes, such as gen_ai.crewai.operation, while GenAI message content capture follows the shared util-genai controls. Message content capture is disabled by default; set both environment variables below before process start to capture gen_ai.input.messages, gen_ai.output.messages, system instructions, and content-like CrewAI fields such as task descriptions or agent backstories:

export OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental
export OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_ONLY

CrewAI has its own cloud tracing path. Set CREWAI_TRACING_ENABLED=false when you only want LoongSuite/OpenTelemetry instrumentation for a process.

Smoke Examples

When running from a source checkout, the examples/crewai_smoke.py script exercises real non-streaming, streaming, and concurrent CrewAI calls. It reads credentials from DASHSCOPE_API_KEY or OPENAI_API_KEY and defaults to DashScope's OpenAI-compatible endpoint. When only DASHSCOPE_API_KEY is set, the example also sets OPENAI_API_KEY, OPENAI_API_BASE, and DASHSCOPE_API_BASE in the current process so CrewAI and LiteLLM can use the OpenAI-compatible DashScope endpoint consistently. Set CREWAI_SMOKE_MODEL=openai/qwen-turbo when validating tool-call paths through LiteLLM's OpenAI-compatible DashScope provider.

loongsuite-instrument python \
  instrumentation-loongsuite/loongsuite-instrumentation-crewai/examples/crewai_smoke.py \
  --mode sync

loongsuite-instrument python \
  instrumentation-loongsuite/loongsuite-instrumentation-crewai/examples/crewai_smoke.py \
  --mode stream

loongsuite-instrument python \
  instrumentation-loongsuite/loongsuite-instrumentation-crewai/examples/crewai_smoke.py \
  --mode concurrent --concurrency 3

For local otel-gui verification, direct OTLP traces to the local backend and let the example configure an HTTP exporter:

export OTEL_SERVICE_NAME=loongsuite-crewai-smoke
export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://127.0.0.1:5173/v1/traces
export CREWAI_SMOKE_MANUAL_INSTRUMENT=true
export CREWAI_SMOKE_CONFIGURE_OTLP=true

python instrumentation-loongsuite/loongsuite-instrumentation-crewai/examples/crewai_smoke.py \
  --mode concurrent --concurrency 3

Supported Versions

• CrewAI >= 0.80.0

References

• CrewAI Documentation
• OpenTelemetry Python Documentation