Skip to content

Langchain-dev-utils Example Project

This repository provides an example project langchain-dev-utils-example designed to help developers quickly understand how to use the utility functions provided by langchain-dev-utils to efficiently build three typical agent architectures:

  • Single Agent: Suitable for executing simple tasks and tasks related to long-term memory storage.
  • Supervisor-Multi-Agent Architecture: Coordinates multiple specialized agents through a central supervisor, suitable for complex scenarios requiring task decomposition, planning, and iterative optimization.
  • Handoffs Architecture: Each agent is responsible for a specific task and can transfer control to other agents when needed.

graph

Quick Start

  1. Clone this repository: 
    git clone https://github.com/TBice123123/langchain-dev-utils-example.git  
cd langchain-dev-utils-example
  2. Install dependencies using uv: 
    uv sync
  3. Create a .env file 
    cp .env.example .env

  4. Edit the .env file and fill in your API keys (requires ZhipuAI and Tavily API keys).

  5. Start the project 

    langgraph dev

Features Used

Single Agent:

Features from this library used:

  • Chat model management (including OpenAI compatible API integration): register_model_provider, load_chat_model
  • Embedding model management: register_embeddings_provider, load_embeddings
  • Format sequence: format_sequence
  • Middleware: format_prompt

Supervisor-Multi-Agent Architecture:

Features from this library used:

  • Chat model management (including OpenAI compatible API integration): register_model_provider, load_chat_model
  • Multi-agent construction: wrap_all_agents_as_tool

Handoffs Architecture:

Features from this library used:

  • Chat model management (including OpenAI compatible API integration): register_model_provider, load_chat_model
  • Middleware: HandoffAgentMiddleware

How to Customize

You can customize this project according to your actual needs.

1. Replace Chat Model Provider

This project uses ZhipuAI's GLM series as the core model by default, as follows:

  • GLM-5: Used as the main agent for simple-agent and supervisor-agent, and for agents in the handoffs architecture
  • GLM-4.7-Flash: Used as the search subagent for supervisor-agent
  • GLM-4.6V: Used as the vision subagent for supervisor-agent

To customize the model provider, please modify src/utils/providers/chat_models/register.py and register your model provider using the register_model_provider function in the register_all_model_providers function.

It is also recommended to modify src/utils/providers/chat_models/load.py and add corresponding loading logic in the load_chat_model function.

Chat Model Management Best Practice

The load_chat_model function uses keyword arguments to receive additional parameters for different chat model classes (LangChain official functions also use this approach). This approach improves universality but weakens IDE type hints and increases the risk of parameter misuse. Therefore, if a specific provider is already determined, you can extend the parameter signature for its integrated chat model class (or embedding model class) to restore type hints. Refer to src/utils/providers/chat_models/load.py for targeted modifications.

2. Register Embedding Model Provider

The registration method for embedding model providers is similar to chat models. Please modify src/utils/providers/embeddings/register.py and register your embedding model provider using the register_embeddings_provider function in the register_all_embeddings_providers function.

To customize the loading logic, you can modify src/utils/providers/embeddings/load.py and add corresponding loading logic in the load_embeddings function.

3. Customize Tools

Single Agent (simple-agent)
Tool implementations are located in src/agents/simple_agent/tools.py, with built-in:
- save_user_memory — Persist user memory
- get_user_memory — Read user memory

To extend, simply add new tool implementations directly in this file.

Supervisor-Multi-Agent (supervisor-agent)
Tool implementations are located in src/agents/supervisor/subagent/tools.py. These are tool implementations for sub-agents. To add custom tools for sub-agents, simply add new tool implementations directly in this file.

Note: The supervisor only has the task tool for calling sub-agents by default. If you need to add custom tools for the supervisor, it is recommended to create a new tools.py under src/agents/supervisor/, write the implementations, and then import and pass them to the create_agent function in src/agents/supervisor/agent.py.

Handoffs Architecture (handoffs)
Tool implementations are located in src/agents/handoffs/tools.py. These are the tool implementations for all agents in this architecture. To add a custom tool for a specific agent, simply add the corresponding tool implementation directly in this file.