framework-packaging-guides

Framework-Specific Packaging Guides

How to turn an agent built with any major framework into something other people can install with one command.

---

1. LangChain / LangGraph

When to use: Stateful multi-step agents with tool calling, memory, and conditional branching. LangGraph adds explicit graph-based control flow.

Project structure:

my-langgraph-agent/
├── src/my_agent/
│   ├── __init__.py
│   ├── agent.py          # graph definition
│   ├── nodes.py          # node functions
│   ├── tools.py
│   └── state.py          # TypedDict state schema
├── langgraph.json        # deployment manifest
├── pyproject.toml
├── Dockerfile
├── docker-compose.yml
└── .env.example

Key configs:

// langgraph.json — required for langgraph deploy and Cloud
{ "graphs": {"my_agent": "./src/my_agent/agent.py:graph"}, "env": ".env", "python_version": "3.12", "dependencies": ["."] }

# pyproject.toml
[project]
name = "my-langgraph-agent"
dependencies = ["langgraph>=0.2.0", "langchain-anthropic>=0.2.0"]
[project.scripts]
my-agent = "my_agent.cli:main"

Distribution: LangGraph Cloud (langgraph deploy) | Docker (docker compose up -d) | PyPI (pip install my-langgraph-agent)

Pitfalls:

1. Version conflicts. LangChain ships breaking changes frequently. Pin exact versions (langgraph==0.2.14, not >=0.2). Users with other LangChain projects hit dependency hell.
2. Missing langgraph.json. Without it, langgraph deploy silently fails or serves the wrong graph.
3. State schema not exported. Library consumers need your State TypedDict. Export it from __init__.py.

---

2. CrewAI

When to use: Multi-agent systems where each agent has a distinct role, goal, and backstory. Role-based collaboration (researcher + writer + editor).

Project structure:

my-crew/
├── src/my_crew/
│   ├── __init__.py
│   ├── crew.py           # @CrewBase class
│   ├── agents.py
│   ├── tasks.py
│   ├── tools/custom_tool.py
│   └── config/
│       ├── agents.yaml
│       └── tasks.yaml
├── pyproject.toml
├── Dockerfile
└── .env.example