Order Processing Agent Workshop

Hands-on tutorials for building, deploying, and evaluating an order operations agent with Deep Agents and LangSmith.

The agent researches payer, coding, and authorization requirements for medical device orders and drafts exception notes the order desk can act on. All order, account, and payer data in these modules is synthetic.

Derived from the Modular Workshops series — this fork keeps the three modules that take an agent from prototype to production. Intended to be run in a session with a LangChain engineer. For more depth on your own, see LangChain Academy, which has pre-recorded videos from our engineers.

Modules

#	Module	Covers	Time
1	01_deep_agents.ipynb	Build the order operations agent on the Deep Agents harness — custom tools, subagents, backends & memory, middleware, HITL, AGENTS.md + skills	~60 min
2	02_deploy.ipynb	Ship that same agent to LangSmith Deployments with the langgraph CLI	~15 min
3	03_langsmith.ipynb	Prompt engineering (Playground + Prompt Hub), tracing, offline & online evals, annotation queues	~30 min

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	Get one
WORKSHOP_USER	All modules — your unique attendee slug	pick one, e.g. jane-doe
OPENAI_API_KEY	Modules 1-3 (default model)	https://platform.openai.com
LANGSMITH_API_KEY	Modules 2 & 3 (recommended for all)	https://smith.langchain.com
TAVILY_API_KEY	Modules 1 & 3 (web search tool)	https://tavily.com
LANGSMITH_API_KEY_GATEWAY	Optional — only if you switch utils/models.py to the LLM Gateway block	same key as LANGSMITH_API_KEY

# 3. Start Jupyter
uv run jupyter notebook

Open whichever module(s) your recipe calls for.

Running with Multiple Attendees

Everyone in a session shares one LangSmith workspace, and LangSmith resources are addressed by name — so two attendees running the same cell would otherwise overwrite each other's tracing project, hub prompt, eval dataset, deployment, run rules, and annotation queue.

WORKSHOP_USER in your .env is the per-attendee slug that keeps them apart. Every shared name is wrapped in scoped() from utils/workshop.py, which suffixes it with your slug:

from utils.workshop import scoped
scoped("order-agent-evals")   # -> "order-agent-evals-jane-doe"

The modules fail loudly at their setup cell if WORKSHOP_USER is unset or still <first-last>, rather than silently colliding.

Use lowercase letters, digits, and hyphens only. .env interpolates the raw value into LANGSMITH_PROJECT, while scoped() slugifies it — so "Jane Doe" would name your tracing project ...-Jane Doe but your dataset ...-jane-doe. You'll get a warning if the two diverge.

Switching Models

All modules import model from utils/models.py. Change one line there to swap providers — no notebook edits required.

# utils/models.py

# OpenAI
# model = init_chat_model("openai:gpt-5.6-terra", use_responses_api=True)

# Anthropic
# model = init_chat_model("anthropic:claude-sonnet-5")

# Azure OpenAI
# from langchain_openai import AzureChatOpenAI
# model = AzureChatOpenAI(azure_deployment="gpt-5.6-terra", streaming=True)

# AWS Bedrock (default)
from langchain_aws import ChatBedrockConverse
model = ChatBedrockConverse(provider="anthropic", model_id="anthropic.claude-sonnet-5")

Routing through the LangSmith LLM Gateway

The simplest route needs no code change — set these in .env and the OpenAI SDK picks them up:

OPENAI_API_KEY="<gateway-key>"
OPENAI_BASE_URL="https://gateway.smith.langchain.com/openai/v1"

The Anthropic equivalent is ANTHROPIC_BASE_URL="https://gateway.smith.langchain.com/anthropic". Both notebooks and the deployed agent then route through the gateway, subject to workspace policies.

utils/models.py also ships a commented-out gateway block that hardcodes the base_url and reads LANGSMITH_API_KEY_GATEWAY (the same key under a non-reserved name, since langgraph deploy strips LANGSMITH_API_KEY during upload). The env-var route above is usually simpler.

The gateway must be enabled for your LangSmith organization, and the key needs the gateway:invoke permission — otherwise calls fail with missing permission gateway:invoke.

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.

Because agents/deep_agent/agent.py imports model from utils.models, whichever block is active in utils/models.py at deploy time is what ships — no extra flags.

Your LANGSMITH_API_KEY must have deployment permissions (use a lsv2_sk_... service key).

Project Structure

order-processing-agent-workshop/
├── README.md                       (this file — recipes + setup)
├── pyproject.toml                  (shared dependencies)
├── .env.example
├── langgraph.json                  (registers agents/deep_agent for langgraph dev)
├── utils/
├── agents/
│   ├── order_agent.py              (shared agent factory — Module 1 builds it, Module 3 imports for eval)
│   └── deep_agent/                 (deployable agent for Module 2)
│       ├── agent.py
│       ├── AGENTS.md
│       └── skills/
│           ├── exception-note/SKILL.md
│           └── status-update/SKILL.md
├── images/                         (diagrams used by the notebooks)
└── modules/
    ├── 01_deep_agents.ipynb        (Module 1 — Deep Agents)
    ├── 02_deploy.ipynb             (Module 2 — Deploy)
    └── 03_langsmith.ipynb          (Module 3 — LangSmith)

Common Issues

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

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

For LangChain Internal Users

Please refer to this linked Notion document for instructions on setup and usage.