Agentscope

中文主页 | Tutorial | Roadmap (Jan 2026 -) | FAQ

What is AgentScope?

AgentScope is a production-ready, easy-to-use agent framework with essential abstractions that work with rising model capability and built-in support for finetuning.

We design for increasingly agentic LLMs. Our approach leverages the models' reasoning and tool use abilities rather than constraining them with strict prompts and opinionated orchestrations.

Why use AgentScope?

• Simple: start building your agents in 5 minutes with built-in ReAct agent, tools, skills, human-in-the-loop steering, memory, planning, realtime voice, evaluation and model finetuning
• Extensible: large number of ecosystem integrations for tools, memory and observability; built-in support for MCP and A2A; message hub for flexible multi-agent orchestration and workflows
• Production-ready: deploy and serve your agents locally, as serverless in the cloud, or on your K8s cluster with built-in OTel support

The AgentScope Ecosystem

News

• [2026-04] COMM: AgentScope 2.0 is on the way. Roadmap | Discussion
• [2026-02] FEAT: Realtime Voice Agent support. Example | Multi-Agent Realtime Example | Tutorial
• [2026-01] COMM: Biweekly Meetings launched to share ecosystem updates and development plans - join us! Details & Schedule
• [2026-01] FEAT: Database support & memory compression in memory module. Example | Tutorial
• [2025-12] INTG: A2A (Agent-to-Agent) protocol support. Example | Tutorial
• [2025-12] FEAT: TTS (Text-to-Speech) support. Example | Tutorial
• [2025-11] INTG: Anthropic Agent Skill support. Example | Tutorial
• [2025-11] RELS: Alias-Agent for diverse real-world tasks and Data-Juicer Agent for data processing open-sourced. Alias-Agent | Data-Juicer Agent
• [2025-11] INTG: Agentic RL via Trinity-RFT library. Example | Trinity-RFT
• [2025-11] INTG: ReMe for enhanced long-term memory. Example

More news →

Community

Welcome to join our community on

Discord	DingTalk

📑 Table of Contents

• Quickstart
  • Installation
    • From PyPI
    • From source
• Example
  • Hello AgentScope!
  • Voice Agent
  • Realtime Voice Agent
  • Human-in-the-loop
  • Flexible MCP Usage
  • Agentic RL
  • Multi-Agent Workflows
• Documentation
• More Examples & Samples
  • Functionality
  • Agent
  • Game
  • Workflow
  • Evaluation
  • Tuner
• Contributing
• License
• Publications
• Contributors

Quickstart

Installation

AgentScope requires Python 3.10 or higher.

From PyPI

pip install agentscope

Or with uv:

uv pip install agentscope

From source

# Pull the source code from GitHub
git clone -b main https://github.com/agentscope-ai/agentscope.git

# Install the package in editable mode
cd agentscope

pip install -e .
# or with uv:
# uv pip install -e .

Example

Hello AgentScope!

Start with a conversation between user and a ReAct agent 🤖 named "Friday"!

from agentscope.agent import ReActAgent, UserAgent
from agentscope.model import DashScopeChatModel
from agentscope.formatter import DashScopeChatFormatter
from agentscope.memory import InMemoryMemory
from agentscope.tool import Toolkit, execute_python_code, execute_shell_command
import os, asyncio


async def main():
    toolkit = Toolkit()
    toolkit.register_tool_function(execute_python_code)
    toolkit.register_tool_function(execute_shell_command)

    agent = ReActAgent(
        name="Friday",
        sys_prompt="You're a helpful assistant named Friday.",
        model=DashScopeChatModel(
            model_name="qwen-max",
            api_key=os.environ["DASHSCOPE_API_KEY"],
            stream=True,
        ),
        memory=InMemoryMemory(),
        formatter=DashScopeChatFormatter(),
        toolkit=toolkit,
    )

    user = UserAgent(name="user")

    msg = None
    while True:
        msg = await agent(msg)
        msg = await user(msg)
        if msg.get_text_content() == "exit":
            break

asyncio.run(main())

Voice Agent

Create a voice-enabled ReAct agent that can understand and respond with speech, even playing a multi-agent werewolf game with voice interactions.

https://github.com/user-attachments/assets/c5f05254-aff6-4375-90df-85e8da95d5da

Realtime Voice Agent

Build a realtime voice agent with web interface that can interact with users via voice input and output.

Realtime chatbot | Realtime Multi-Agent Example

https://github.com/user-attachments/assets/1b7b114b-e995-4586-9b3f-d3bb9fcd2558

Human-in-the-loop

Support realtime interruption in ReActAgent: conversation can be interrupted via cancellation in realtime and resumed seamlessly via robust memory preservation.

Flexible MCP Usage

Use individual MCP tools as local callable functions to compose toolkits or wrap into a more complex tool.

from agentscope.mcp import HttpStatelessClient
from agentscope.tool import Toolkit
import os

async def fine_grained_mcp_control():
    # Initialize the MCP client
    client = HttpStatelessClient(
        name="gaode_mcp",
        transport="streamable_http",
        url=f"https://mcp.amap.com/mcp?key={os.environ['GAODE_API_KEY']}",
    )

    # Obtain the MCP tool as a **local callable function**, and use it anywhere
    func = await client.get_callable_function(func_name="maps_geo")

    # Option 1: Call directly
    await func(address="Tiananmen Square", city="Beijing")

    # Option 2: Pass to agent as a tool
    toolkit = Toolkit()
    toolkit.register_tool_function(func)
    # ...

    # Option 3: Wrap into a more complex tool
    # ...

Agentic RL

Train your agentic application seamlessly with Reinforcement Learning integration. We also prepare multiple sample projects covering various scenarios:

Example	Description	Model	Training Result
Math Agent	Tune a math-solving agent with multi-step reasoning.	Qwen3-0.6B	Accuracy: 75% → 85%
Frozen Lake	Train an agent to navigate the Frozen Lake environment.	Qwen2.5-3B-Instruct	Success rate: 15% → 86%
Learn to Ask	Tune agents using LLM-as-a-judge for automated feedback.	Qwen2.5-7B-Instruct	Accuracy: 47% → 92%
Email Search	Improve tool-use capabilities without labeled ground truth.	Qwen3-4B-Instruct-2507	Accuracy: 60%
Werewolf Game	Train agents for strategic multi-agent game interactions.	Qwen2.5-7B-Instruct	Werewolf win rate: 50% → 80%
Data Augment	Generate synthetic training data to enhance tuning results.	Qwen3-0.6B	AIME-24 accuracy: 20% → 60%

Multi-Agent Workflows

AgentScope provides MsgHub and pipelines to streamline multi-agent conversations, offering efficient message routing and seamless information sharing

from agentscope.pipeline import MsgHub, sequential_pipeline
from agentscope.message import Msg
import asyncio

async def multi_agent_conversation():
    # Create agents
    agent1 = ...
    agent2 = ...
    agent3 = ...
    agent4 = ...

    # Create a message hub to manage multi-agent conversation
    async with MsgHub(
        participants=[agent1, agent2, agent3],
        announcement=Msg("Host", "Introduce yourselves.", "assistant")
    ) as hub:
        # Speak in a sequential manner
        await sequential_pipeline([agent1, agent2, agent3])
        # Dynamic manage the participants
        hub.add(agent4)
        hub.delete(agent3)
        await hub.broadcast(Msg("Host", "Goodbye!", "assistant"))

asyncio.run(multi_agent_conversation())

Documentation

• Tutorial
• FAQ
• API Docs

More Examples & Samples

Functionality

• MCP
• Anthropic Agent Skill
• Plan
• Structured Output
• RAG
• Long-Term Memory
• Session with SQLite
• Stream Printing Messages
• TTS
• Code-first Deployment
• Memory Compression

Agent

• ReAct Agent
• Voice Agent
• Deep Research Agent
• Browser-use Agent
• Meta Planner Agent
• A2A Agent
• Realtime Voice Agent

Game

• Nine-player Werewolves

Workflow

• Multi-agent Debate
• Multi-agent Conversation
• Multi-agent Concurrent
• Multi-agent Realtime Conversation

Evaluation

• ACEBench

Tuner

• Tune ReAct Agent

Contributing

We welcome contributions from the community! Please refer to our CONTRIBUTING.md for guidelines on how to contribute.

License

AgentScope is released under Apache License 2.0.

Publications

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

• AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications

• AgentScope: A Flexible yet Robust Multi-Agent Platform

@article{agentscope_v1,
    author  = {Dawei Gao, Zitao Li, Yuexiang Xie, Weirui Kuang, Liuyi Yao, Bingchen Qian, Zhijian Ma, Yue Cui, Haohao Luo, Shen Li, Lu Yi, Yi Yu, Shiqi He, Zhiling Luo, Wenmeng Zhou, Zhicheng Zhang, Xuguang He, Ziqian Chen, Weikai Liao, Farruh Isakulovich Kushnazarov, Yaliang Li, Bolin Ding, Jingren Zhou}
    title   = {AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications},
    journal = {CoRR},
    volume  = {abs/2508.16279},
    year    = {2025},
}

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

Contributors

All thanks to our contributors: