Framework-agnostic LangSmith tracing demo

A grocery in-store shopping assistant built three different ways — to show that LangSmith tracing is framework-agnostic. All three implementations route their model calls through LangChain's LLM gateway and trace to LangSmith:

Implementation	Framework	Model (via gateway)	Tracing
agent_adk.py	Google ADK	gemini-3.5-flash-lite (LiteLlm)	native configure_google_adk
agent_openai_sdk.py	OpenAI Agents SDK	gpt-5.6-luna	OpenAIAgentsTracingProcessor
agent_bare.py	none (hand-written tool loop)	gpt-5.6-luna	wrap_openai + @traceable

All three share the same in-store shopping tools (grocery_tools.py): find an item's aisle, check price, check stock, and suggest a recipe shopping list.

The LangChain LLM gateway is the single point every implementation calls:

• Gemini endpoint: https://gateway.smith.langchain.com/gemini
• OpenAI-compatible endpoint: https://gateway.smith.langchain.com/openai/v1

Your LANGSMITH_API_KEY is used both as the gateway API key and to send traces to LangSmith.

Setup

Requires Python 3.11+ and uv.

uv sync

Copy .env.example to .env and fill in your LangSmith API key:

LANGSMITH_API_KEY=<your key>
LANGSMITH_PROJECT="grocery-agents"
LANGSMITH_TRACING="true"

Run the UI (recommended)

A Streamlit chat UI lets you pick which implementation to talk to:

uv run streamlit run app.py

Use the sidebar to switch between Google ADK, OpenAI Agents SDK, and Bare code, then chat. Open your LangSmith project to watch traces land from all three.

Run from the CLI

Fire one question at each of the three implementations in sequence:

uv run main.py

Traces from every implementation show up in the LangSmith project set by LANGSMITH_PROJECT.