How LoongSuite Python Probe Simplifies AI Agent Observability

Observability challenges for AI agents

When AI applications involve multiple agents, tool calls, retrieval‑augmented generation (RAG), and memory, developers lose visibility into latency sources, context changes, and token costs. Traditional micro‑service observability focuses on performance and availability, but AI workloads require tracing of dialogues, tool invocations, retrieval results, and multimodal inputs.

Lightweight capture of context changes across frameworks and business code.

Efficient handling of large multimodal payloads (images, audio, video) without slowing the system.

Unified semantic model for observability data across different tools.

LoongSuite Python Probe

LoongSuite Python Probe is Alibaba Cloud’s open‑source distribution of OpenTelemetry Python Contrib. It provides automatic instrumentation for popular AI frameworks such as DashScope, LangChain, AgentScope, Dify, MCP, Mem0, and others. The probe records the model name, tool name, token usage, and context changes, and emits spans and events that follow the OpenTelemetry GenAI semantic conventions, making the data consumable by any OTLP‑compatible backend (Jaeger, Langfuse, Alibaba Cloud Observability, etc.).

How the probe works

Automatic discovery : at startup the probe scans installed packages and loads the corresponding instrumentation modules.

Unified semantics : spans and events use the GenAI attribute names defined by the OpenTelemetry GenAI SIG.

Multi‑dimensional coverage : traces cover LLM calls, agent actions, tool invocations, HTTP/gRPC/database calls, providing end‑to‑end visibility.

Flexible export : data are sent via the OTLP exporter to any backend that supports the OTLP protocol.

Quick‑start (three steps)

pip install loongsuite-distro

loongsuite-bootstrap -a install --version 0.1.0

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317 \
OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental \
OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=SPAN_ONLY \
loongsuite-instrument python app.py

After these commands the AI application is fully observable.

Manual instrumentation with LoongSuite GenAI Util

For code paths that the automatic probe cannot detect, developers can use the GenAI Util. Example:

from opentelemetry.util.genai.extended_handler import get_extended_telemetry_handler
from opentelemetry.util.genai.extended_types import InvokeAgentInvocation
from opentelemetry.util.genai.types import InputMessage, OutputMessage, Text

handler = get_extended_telemetry_handler()
invocation = InvokeAgentInvocation(
    provider="dashscope",
    request_model="qwen-turbo",
    agent_name="OrderAgent",
    input_messages=[
        InputMessage(role="user", parts=[Text(content="查询订单 101 状态")]),
        InputMessage(role="system", parts=[Text(content="你是订单管理员，调用工具查询信息")])
    ]
)
with handler.invoke_agent(invocation) as inv:
    # invoke the agent …
    inv.output_messages = [
        OutputMessage(
            role="assistant",
            parts=[Text(content="订单未找到，请检查单号。")],
            finish_reason="stop"
        )
    ]
    inv.input_tokens = 15
    inv.output_tokens = 20

The util also provides asynchronous multimodal upload pipelines (PreUploader and Uploader) that replace large blobs with URIs in the trace, preventing payloads from blocking the request path.

Release notes and roadmap

The probe is released as the loongsuite-distro package (bootstrap and instrument commands) and the loongsuite-util-genai library (enhanced GenAI utilities). Future work includes expanding the plugin matrix for domestic frameworks, adding more span types (invoke_agent, create_agent, execute_tool, retrieve, rerank, embedding, memory), and improving multimodal handling. Full release notes: https://github.com/alibaba/loongsuite-python-agent/releases

Key advantages

Zero‑code‑change, OpenTelemetry‑native observability for AI agents.

Unified GenAI semantics enable seamless integration with Jaeger, Langfuse, Alibaba Cloud Observability, and other OTLP backends.

Support for multimodal payload off‑loading to OSS, SLS, or local storage via the GenAI Util.