Langchain Agents Scaffold

Scaffolding LangChain ecosystem projects

There is no single scaffolder that covers all three project shapes. Pick the right path:

LangGraph: langgraph new

pip install "langgraph-cli>=0.4"    # if not installed
langgraph new my-agent --template react-agent
cd my-agent
pip install -e .

langgraph-cli ships several templates. List them with langgraph new --help. Common picks:

• react-agent — single-LLM-with-tools loop. The most common starting point.
• retrieval-agent — RAG over a vector store.
• memory-agent — long-term memory using the LangGraph store.
• data-enrichment-agent — structured data extraction.

Each template ships its own pyproject.toml, langgraph.json, and src/<package>/graph.py — read those after scaffolding to learn the layout. Do not assume the layout matches across templates. The conventions vary.

DeepAgents: write the file directly

There's no deepagents new. Create the project by hand:

mkdir my-deep-agent && cd my-deep-agent
python -m venv .venv && source .venv/bin/activate
pip install \
  "deepagents>=0.5.3" \
  "langchain>=1.2" \
  "langchain-anthropic>=1.4" \
  "langsmith>=0.7"
mkdir agent

Pin floors matter: deepagents>=0.5 removed the legacy subagents API and added async sub-agents; 0.5.2 added the filesystem permissions system; 0.5.3 made model=None a deprecated kwarg (you must pass an explicit model) and added structured outputs for sub-agent responses.

Then agent/__init__.py (empty) and agent/agent.py:

"""DeepAgent for my-deep-agent. Always exported as `agent`."""
from deepagents import create_deep_agent
from langchain.chat_models import init_chat_model

SYSTEM_PROMPT = "You are my-deep-agent, a helpful agent."

TOOLS = []          # add user tools here
SUBAGENTS = []      # add sub-agents here (see deepagents-code skill)

agent = create_deep_agent(
    model=init_chat_model("anthropic:claude-sonnet-4-6"),  # explicit model required as of 0.5.3
    tools=TOOLS,
    subagents=SUBAGENTS,
    instructions=SYSTEM_PROMPT,
)

Plus a pyproject.toml (or requirements.txt) and a .env with ANTHROPIC_API_KEY and LANGSMITH_*. That's the whole project.

For deploy: DeepAgents' create_deep_agent returns a compiled LangGraph, so a langgraph.json pointing at agent.agent:agent works for langgraph dev and langgraph build/deploy.

LCEL chains: write the file directly

For non-agentic flows (RAG, summarization, classification):

mkdir my-chain && cd my-chain
python -m venv .venv && source .venv/bin/activate
pip install "langchain>=1.2" "langchain-openai>=1.0" "langsmith>=0.7"
mkdir agent

Then agent/agent.py:

"""LCEL chain. Exposed as `agent` (a Runnable)."""
from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_core.runnables import RunnableLambda

SYSTEM_PROMPT = "You are a helpful assistant."

def _to_messages(payload: dict) -> list:
    msgs = [SystemMessage(content=SYSTEM_PROMPT)]
    for m in payload.get("messages", []):
        msgs.append(HumanMessage(content=m["content"]))
    return msgs

agent = RunnableLambda(_to_messages) | init_chat_model("openai:gpt-4o-mini")

For RAG, see the langchain-agents-langchain-code skill.

Naming conventions worth following (not enforced)

These are conventions, not requirements. They make follow-up work easier because every other skill in this bundle assumes them:

• The runnable artifact is named agent and lives at agent/agent.py.
• Provider keys and LANGSMITH_* go in .env. Commit a .env.example.
• Evalsets live under evals/datasets/*.jsonl; the eval runner at evals/run.py.
• A FastAPI host (if needed for Docker/Cloud Run deploy) lives at server/app.py.

Skills that follow assume these names. If the project diverges, adapt — these are not hard rules, just the path of least resistance.