Deep Agent Assistants

A walkthrough of deploying a configurable Deep Agent to LangSmith Deployments and using the LangGraph SDK to create and manage multiple assistant configurations on top of a single deployed graph.

Overview

This repo demonstrates a core LangSmith pattern: one deployed graph, many assistants. The deep_agent graph exposes configurable parameters (model, system prompt, tools) so you can create distinct assistant personas — each with their own behavior, tool access, and long-term memory — without redeploying.

Prerequisites

• Python 3.11+
• A LangSmith account with API access
• An Anthropic API key (for Claude models)
• A Tavily API key (for research tools)

Environment Variables

Create a .env file in the root of the repo:

LANGSMITH_API_KEY=your_langsmith_api_key
ANTHROPIC_API_KEY=your_anthropic_api_key
TAVILY_API_KEY=your_tavily_api_key

Deployment

You have two options to run the agent:

Option 1 — Local development with LangGraph Dev

uv sync
uv run langgraph dev

This starts a local LangGraph server at http://localhost:2024. Use this URL in the notebook client setup.

Option 2 — LangSmith Cloud Deployment

1. Push this repo to GitHub
2. In LangSmith, go to Deployments and create a new deployment
3. Connect your GitHub repo — LangSmith will build and deploy the graph automatically
4. Copy the deployment URL from the LangSmith UI

The langgraph.json file tells LangSmith which graphs to deploy:

{
  "graphs": {
    "agent": "./src/deep_research/agent.py:agent",
    "deep_agent": "./src/deep_agent/agent.py:make_graph"
  }
}

Notebook Walkthrough

Once the agent is deployed (locally or in the cloud), open notebooks/assistant_test.ipynb and work through the cells in order:

1. Connect to the deployed graph

Update the url in the client setup cell to point to your deployment (local or cloud):

client = get_client(
    url="https://your-deployment-url.us.langgraph.app",
    api_key=os.environ["LANGSMITH_API_KEY"],
)

2. Create assistants

Each assistant is a named configuration on top of the deep_agent graph. You can customize:

• system_prompt — the assistant's persona and behavior
• model — the LLM to use (e.g. anthropic:claude-haiku-4-5)
• selected_tools — which tools the assistant has access to
• name — display name for the assistant

assistant = await client.assistants.create(
    graph_id="deep_agent",
    config={
        "configurable": {
            "system_prompt": "You are a helpful AI assistant. Always reply in the voice of a pirate.",
            "model": "anthropic:claude-haiku-4-5",
            "selected_tools": ["get_todays_date", "advanced_research"],
            "name": "Pirate"
        }
    },
    name="Pirate",
)

Available tools:

Tool	Description
get_todays_date	Returns the current date
advanced_research	Deep web research via Tavily (10 results)
basic_research	Quick web search via Tavily (5 results)
finance_research	Yahoo Finance news by ticker symbol
langchain_docs	Search LangChain documentation via MCP

3. Seed assistant memory

Each assistant can have a persistent AGENTS.md file stored in the LangGraph store. This file is loaded at startup and scopes the assistant's long-term memory to its assistant_id. The notebook seeds each assistant's memory from files in src/assistants/:

src/assistants/
├── content-writer/AGENTS.md   # Persona instructions for the Pirate assistant
└── mcp-docs-agent/AGENTS.md   # Persona instructions for the Cowboy assistant

4. Stream a response

The final cells create a thread, send a message to an assistant, and stream the response token-by-token using stream_mode="messages".

Project Structure

.
├── langgraph.json                  # Graph entrypoints for deployment
├── pyproject.toml                  # Dependencies
├── notebooks/
│   └── assistant_test.ipynb        # Main walkthrough notebook
└── src/
    ├── assistants/
    │   ├── content-writer/AGENTS.md
    │   └── mcp-docs-agent/AGENTS.md
    └── deep_agent/
        ├── agent.py                # Graph definition (make_graph)
        ├── context.py              # Configurable assistant parameters
        └── tools.py                # Custom tools