Finance Research Agent Workshop

A guided tour of the LangChain platform built around a single use case: a financial research agent that goes from sandbox to a governed production deployment with full observability. Four standalone Jupyter modules that combine into ~1.5 to 2 hour workshops.

The agent itself is a generalist financial research analyst — it researches equities, fixed income, macro, and sectors from public sources, then writes investor notes or earnings summaries. The same agent serves as the running example across every module.

The Modules

#	Module	Duration	Notebook
1	Deep Agents — harness, custom tools, subagents, memory, middleware, HITL, AGENTS.md and Skills	~45 min	modules/01_deep_agents.ipynb
2	Move to Production — workspace policies via the LangSmith LLM Gateway, then deploy with langgraph CLI to LangSmith Deployments	~25 min	modules/02_deploy_and_govern.ipynb
3	LangSmith — tracing and querying traces, Engine for automated failure-mode discovery, offline + online evaluations, annotation queues	~30 min	modules/03_langsmith.ipynb
4	Accelerate LangGraph with NVIDIA — parallel and speculative execution via langchain-nvidia-langgraph	~10 min	modules/04_nvidia.ipynb

Each module is a standalone Jupyter notebook. Modules share the project's setup, utils/, and agents/, so combining them is as straightforward as opening multiple notebooks in order.

Prerequisites

• Python 3.11+
• uv (recommended) or pip

Setup

# 1. Install dependencies
uv sync

# 2. Configure environment variables
cp .env.example .env
# Edit .env and fill in your keys

Key	Required for	Notes
LANGSMITH_API_KEY	All modules	Tracing, deployments, gateway, evaluations. Use a service key (lsv2_sk_...) for Module 2 deploys.
ANTHROPIC_API_KEY	All modules	Default model (claude-sonnet-4-6) is routed through the LangSmith LLM Gateway.
WORKSPACE_ID	Module 2	The LangSmith workspace the gateway policy applies to. Find it in Settings → Workspace.
LANGSMITH_API_KEY_GATEWAY	Module 2 (after §1.4 flip) and the deployed agent	Same value as LANGSMITH_API_KEY. Required under a non-reserved name because langgraph deploy strips LANGSMITH_API_KEY from deployed containers.
TAVILY_API_KEY	Modules 1 and 3	Web search tool used by the research agent. https://tavily.com
OPENAI_API_KEY	Optional	Only required if you swap the default model in utils/models.py to OpenAI.

# 3. Start Jupyter
uv run jupyter notebook

Open whichever module(s) you intend to run.

Default Model and the LangSmith LLM Gateway

The workshop's default model is configured in utils/models.py:

model = init_chat_model(
    model="claude-sonnet-4-6",
    model_provider="anthropic",
    base_url="https://gateway.smith.langchain.com/anthropic",
)

The base_url routes every model call through the LangSmith LLM Gateway, which means any workspace-level policy (PII detection, secrets redaction, allow-lists, rate limits, cost caps) applies uniformly across the workshop — including to the deployable agent. Module 2 walks through configuring a sample policy.

To switch providers, edit utils/models.py:

# Anthropic, direct (bypasses the gateway and any workspace policies)
# model = init_chat_model("anthropic:claude-sonnet-4-6")

# OpenAI
# model = init_chat_model("openai:gpt-4.1-mini")

# Azure OpenAI
# from langchain_openai import AzureChatOpenAI
# model = AzureChatOpenAI(azure_deployment="gpt-4.1-mini", streaming=True)

# AWS Bedrock
# from langchain_aws import ChatBedrockConverse
# model = ChatBedrockConverse(provider="anthropic", model_id="...")

Deploy (Module 2)

Module 2 deploys the agent at agents/deep_agent/ to LangSmith via the langgraph CLI (installed by uv sync). The deploy config is langgraph.json at the workshop root. The deployment name is financial-research-agent.

The LANGSMITH_API_KEY used for deploys must have deployment permissions (a lsv2_sk_... service key, not a personal token).

Project Structure

finance-research-agent-workshop/
├── README.md                       (this file — setup + reference)
├── pyproject.toml                  (shared dependencies)
├── .env.example
├── langgraph.json                  (registers agents/deep_agent for langgraph dev)
├── utils/
│   ├── models.py                   (default model, gateway-routed)
│   ├── search.py                   (Tavily wrapper with topic-matched fallbacks)
│   └── langsmith_rules.py          (run rule + annotation queue helpers)
├── agents/
│   ├── research_agent.py           (shared agent factory — Module 1 references, Module 3 imports for eval)
│   └── deep_agent/                 (deployable agent for Module 2)
│       ├── agent.py
│       ├── AGENTS.md               (agent identity)
│       ├── deepagents.toml
│       └── skills/
│           ├── investor-note/SKILL.md
│           └── earnings-summary/SKILL.md
├── images/                         (diagrams + screenshots used by the notebooks)
└── modules/
    ├── 01_deep_agents.ipynb
    ├── 02_deploy_and_govern.ipynb
    ├── 03_langsmith.ipynb
    └── 04_nvidia.ipynb

Common Issues

langgraph deploy fails with 403 / permission denied The API key in use is a personal token. Generate a service key (lsv2_sk_...) in LangSmith settings.

Module 2 policy creation fails with 401 / unauthorized LANGSMITH_API_KEY lacks the workspace admin scope required to create gateway policies, or WORKSPACE_ID does not match the workspace the key belongs to.

Notebook can't find utils / agents Each module's setup cell prepends the workshop root to sys.path. If a notebook has been moved, update the Path().resolve().parent line to point at the workshop root.