English | 中文

AgentScope

Start building LLM-empowered multi-agent applications in an easier way.

• If you find our work helpful, please kindly cite our paper.

• Visit our workstation to build multi-agent applications with dragging-and-dropping.

• Welcome to join our community on

Discord	DingTalk

News

• [2024-09-06] AgentScope version 0.1.0 is released now.

• [2024-09-03] AgentScope supports Web Browser Control now! Refer to our example for more details.

• [2024-07-18] AgentScope supports streaming mode now! Refer to our tutorial and example conversation in stream mode for more details.

• [2024-07-15] AgentScope has implemented the Mixture-of-Agents algorithm. Refer to our MoA example for more details.

• [2024-06-14] A new prompt tuning module is available in AgentScope to help developers generate and optimize the agents' system prompts! Refer to our tutorial for more details!

• [2024-06-11] The RAG functionality is available for agents in AgentScope now! A quick introduction to RAG in AgentScope can help you equip your agent with external knowledge!

• [2024-06-09] We release AgentScope v0.0.5 now! In this new version, AgentScope Workstation (the online version is running on agentscope.io) is open-sourced with the refactored AgentScope Studio!

What's AgentScope?

AgentScope is an innovative multi-agent platform designed to empower developers to build multi-agent applications with large-scale models. It features three high-level capabilities:

• 🤝 Easy-to-Use: Designed for developers, with fruitful components, comprehensive documentation, and broad compatibility. Besides, AgentScope Workstation provides a drag-and-drop programming platform and a copilot for beginners of AgentScope!

• ✅ High Robustness: Supporting customized fault-tolerance controls and retry mechanisms to enhance application stability.

• 🚀 Actor-Based Distribution: Building distributed multi-agent applications in a centralized programming manner for streamlined development.

Supported Model Libraries

AgentScope provides a list of ModelWrapper to support both local model services and third-party model APIs.

Supported Local Model Deployment

AgentScope enables developers to rapidly deploy local model services using the following libraries.

• ollama (CPU inference)
• Flask + Transformers
• Flask + ModelScope
• FastChat
• vllm

Supported Services

• Web Search
• Data Query
• Retrieval
• Code Execution
• File Operation
• Text Processing
• Multi Modality
• Wikipedia Search and Retrieval
• TripAdvisor Search
• Web Browser Control

Example Applications

• Model

  • Using Llama3 in AgentScope

• Conversation

  • Basic Conversation
  • Autonomous Conversation with Mentions
  • Self-Organizing Conversation
  • Basic Conversation with LangChain library
  • Conversation with ReAct Agent
  • Conversation in Natural Language to Query SQL
  • Conversation with RAG Agent
  • Conversation with gpt-4o
  • Conversation with Software Engineering Agent
  • Conversation with Customized Tools
  • Mixture of Agents Algorithm
  • Conversation in Stream Mode
  • Conversation with CodeAct Agent
  • Conversation with Router Agent

• Game

  • Gomoku
  • Werewolf

• Distribution

  • Distributed Conversation
  • Distributed Debate
  • Distributed Parallel Optimization
  • Distributed Large Scale Simulation

More models, services and examples are coming soon!

Installation

AgentScope requires Python 3.9 or higher.

Note: This project is currently in active development, it's recommended to install AgentScope from source.

From source

• Install AgentScope in editable mode:

# Pull the source code from GitHub
git clone https://github.com/modelscope/agentscope.git

# Install the package in editable mode
cd agentscope
pip install -e .

Using pip

• Install AgentScope from pip:

pip install agentscope

Extra Dependencies

To support different deployment scenarios, AgentScope provides several optional dependencies. Full list of optional dependencies refers to tutorial Taking distribution mode as an example, you can install its dependencies as follows:

On Windows

# From source
pip install -e .[distribute]
# From pypi
pip install agentscope[distribute]

On Mac & Linux

# From source
pip install -e .\[distribute\]
# From pypi
pip install agentscope\[distribute\]

Quick Start

Configuration

In AgentScope, the model deployment and invocation are decoupled by ModelWrapper.

To use these model wrappers, you need to prepare a model config file as follows.

model_config = {
    # The identifies of your config and used model wrapper
    "config_name": "{your_config_name}",          # The name to identify the config
    "model_type": "{model_type}",                 # The type to identify the model wrapper

    # Detailed parameters into initialize the model wrapper
    # ...
}

Taking OpenAI Chat API as an example, the model configuration is as follows:

openai_model_config = {
    "config_name": "my_openai_config",             # The name to identify the config
    "model_type": "openai_chat",                   # The type to identify the model wrapper

    # Detailed parameters into initialize the model wrapper
    "model_name": "gpt-4",                         # The used model in openai API, e.g. gpt-4, gpt-3.5-turbo, etc.
    "api_key": "xxx",                              # The API key for OpenAI API. If not set, env
                                                   # variable OPENAI_API_KEY will be used.
    "organization": "xxx",                         # The organization for OpenAI API. If not set, env
                                                   # variable OPENAI_ORGANIZATION will be used.
}

More details about how to set up local model services and prepare model configurations is in our tutorial.

Create Agents

Create built-in user and assistant agents as follows.

from agentscope.agents import DialogAgent, UserAgent
import agentscope

# Load model configs
agentscope.init(model_configs="./model_configs.json")

# Create a dialog agent and a user agent
dialog_agent = DialogAgent(name="assistant",
                           model_config_name="my_openai_config")
user_agent = UserAgent()

Construct Conversation

In AgentScope, message is the bridge among agents, which is a dict that contains two necessary fields name and content and an optional field url to local files (image, video or audio) or website.

from agentscope.message import Msg

x = Msg(name="Alice", content="Hi!")
x = Msg("Bob", "What about this picture I took?", url="/path/to/picture.jpg")

Start a conversation between two agents (e.g. dialog_agent and user_agent) with the following code:

x = None
while True:
    x = dialog_agent(x)
    x = user_agent(x)
    if x.content == "exit":  # user input "exit" to exit the conversation_basic
        break

AgentScope Studio

AgentScope provides an easy-to-use runtime user interface capable of displaying multimodal output on the front end, including text, images, audio and video.

Refer to our tutorial for more details.

Tutorial

• About AgentScope
• Installation
• Quick Start
• Model
• Prompt Engineering
• Agent
• Memory
• Response Parser
• Tool
• Pipeline and MsgHub
• Distribution
• AgentScope Studio
• Logging
• Monitor
• Example: Werewolf Game

License

AgentScope is released under Apache License 2.0.

Contributing

Contributions are always welcomed!

We provide a developer version with additional pre-commit hooks to perform checks compared to the official version:

# For windows
pip install -e .[dev]
# For mac
pip install -e .\[dev\]

# Install pre-commit hooks
pre-commit install

Please refer to our Contribution Guide for more details.

Publications

If you find our work helpful for your research or application, please cite our papers.

1. AgentScope: A Flexible yet Robust Multi-Agent Platform

   @article{agentscope,
       author  = {Dawei Gao and
                  Zitao Li and
                  Xuchen Pan and
                  Weirui Kuang and
                  Zhijian Ma and
                  Bingchen Qian and
                  Fei Wei and
                  Wenhao Zhang and
                  Yuexiang Xie and
                  Daoyuan Chen and
                  Liuyi Yao and
                  Hongyi Peng and
                  Ze Yu Zhang and
                  Lin Zhu and
                  Chen Cheng and
                  Hongzhu Shi and
                  Yaliang Li and
                  Bolin Ding and
                  Jingren Zhou}
       title   = {AgentScope: A Flexible yet Robust Multi-Agent Platform},
       journal = {CoRR},
       volume  = {abs/2402.14034},
       year    = {2024},
   }

2. On the Design and Analysis of LLM-Based Algorithms

   @article{llm_based_algorithms,
       author  = {Yanxi Chen and
                  Yaliang Li and
                  Bolin Ding and
                  Jingren Zhou},
       title   = {On the Design and Analysis of LLM-Based Algorithms},
       journal = {CoRR},
       volume  = {abs/2407.14788},
       year    = {2024},
   }

3. Very Large-Scale Multi-Agent Simulation in AgentScope

   @article{agentscope_simulation,
       author  = {Xuchen Pan and
                  Dawei Gao and
                  Yuexiang Xie and
                  Zhewei Wei and
                  Yaliang Li and
                  Bolin Ding and
                  Ji{-}Rong Wen and
                  Jingren Zhou},
       title   = {Very Large-Scale Multi-Agent Simulation in AgentScope},
       journal = {CoRR},
       volume  = {abs/2407.17789},
       year    = {2024},
   }