Tracing for Strands Agents

This project demonstrates how to instrument and trace an example strands agent to LangSmith using OpenTelemetry, enabling you to monitor model & agent performance, latency, and token usage.

🛠 Setup Clone the repo

Clone the repo

git clone https://github.com/langchain-ai/strands-otel-tracing-example

Create an environment variables file

$ cd strands-example
# Copy the .env.example file to .env
cp .env.example .env

Fill in fields such as OTel endpoint, headers (project and API key), and AWS credentials

Package Installation

Ensure you have a recent version of uv installed

$ uv sync

Run the agent

$ uv run otel_strands_share.py

You can then see an example trace in the LangSmith project specified!

Using in your own project

Copy langsmith_exporter.py into your project, then call setup_langsmith_telemetry() before creating your agent:

from langsmith_exporter import setup_langsmith_telemetry
from strands import Agent

setup_langsmith_telemetry()

This replaces the standard StrandsTelemetry().setup_otlp_exporter() call. It wraps the OTLP exporter with a transformation layer that:

• Standardizes message attributes — Strands emits messages as span events, but the GenAI semantic conventions specify them as gen_ai.prompt / gen_ai.completion span attributes. The exporter normalizes to the expected format.
• Standardizes content blocks — Converts Bedrock/Converse-shaped blocks ({"text": "..."}, {"toolUse": {...}}) into typed blocks ({"type": "text", "text": "..."}, {"type": "tool_use", ...}).
• Maps run types — Sets langsmith.span.kind based on gen_ai.operation.name so spans render as the correct type in LangSmith (chain for agent invocations, llm for model calls, tool for tool executions).

The exporter reads endpoint and auth configuration from the standard OTEL_EXPORTER_OTLP_* environment variables in this repo's .env.example file.