agentscope-skill（Agent 范围技能）

Understanding AgentScope

What is AgentScope?

AgentScope is a production-ready, enterprise-grade open-source framework for building multi-agent applications with large language models. Its functionalities cover:

• Development: ReAct agent, context compression, short/long-term memory, tool use, human-in-the-loop, multi-agent orchestration, agent hooks, structured output, planning, integration with MCP, agent skill, LLMs API, voice interaction (TTS/Realtime), RAG
• Evaluation: Evaluate multistep agentic applications with statistical analysis
• Training: Agentic reinforcement learning
• Deployment: Session/state management, sandbox, local/serverless/Kubernetes deployment

Installation

pip install agentscope
# or
uv pip install agentscope

Core Concepts

• Message: The core abstraction for information exchange between agents. Supports heterogeneous content blocks (text, images, tool calls, tool results).

from agentscope.message import Msg, TextBlock, ImageBlock, URLSource

msg = Msg(
    name="user",
    content=[TextBlock("Hello world"), ImageBlock(type="image", source=URLSource(type="url", url="..."))],
    role="user"
)

• Agent: LLM-empowered agent that can reason, use tools, and generate responses through iterative thinking and action loops.
• Toolkit: Register and manage tools (Python functions, MCP, agent skills) that agents can call.
• Memory: Store Msg objects as conversation history/context with a marking mechanism for advanced memory management (compression, retrieval).
• ChatModel: Unified interface across different providers (OpenAI, Anthropic, DashScope, Ollama, etc.) with support for tool use and streaming.
• Formatter: Convert Msg objects to LLM API-specific formats. Must be used with the corresponding ChatModel. Supports multi-agent conversations with different agent identifiers.

Basic Usage Examples

Example 1: Simple Chatbot

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():
    # Initialize toolkit with tools
    toolkit = Toolkit()
    toolkit.register_tool_function(execute_python_code)
    toolkit.register_tool_function(execute_shell_command)

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

    # Create user agent for terminal input
    user = UserAgent(name="user")

    # Conversation loop
    msg = None
    while True:
        msg = await agent(msg)  # Agent processes and replies
        msg = await user(msg)   # User inputs next message
        if msg.get_text_content() == "exit":
            break

asyncio.run(main())

Example 2: Multi-Agent Conversation

AgentScope adopts explicit message passing for multi-agent conversations (PyTorch-like dynamic graph), allowing flexible information flow control.

alice, bob, carol, david = ReActAgent(...), ReActAgent(...), ReActAgent(...), ReActAgent(...)

msg_alice = await alice()
msg_bob = await bob(msg_alice)  # Bob receives Alice's message and generate a reply. Alice doesn't receive Bob's message unless explicitly passed back.
msg_carol = await carol(msg_alice)  # Similarly, the agent cannot receive messages from other agents unless explicitly passed.

# Broadcasting with MsgHub, a syntactic sugar for message broadcasting within a group of agents
from agentscope.pipeline import MsgHub

async with MsgHub(
    participants=[alice, bob, carol],
    announcement=Msg("Host", "Let's discuss", "user")
) as hub:
    await alice()  # Bob and Carol receive this
    await bob()    # Alice and Carol receive this

    # Manual broadcast
    await hub.broadcast(Msg("Host", "New topic", "user"))

    # Dynamic participant management
    hub.add(david)
    hub.delete(bob)

Example 3: Master-Worker Pattern

Wrap worker agents as tools for the master agent.

from agentscope.tool import ToolResponse, Toolkit

async def create_worker(task: str) -> ToolResponse:
    """Create a worker agent for the given task.

    Args:
        task (`str`): The given task, which should be specific and concise.
    """
    task_msg = Msg(name="master", content=task, role="user") # Use the input task or wrap it into a more complex prompt
    worker = ReActAgent(...)
    res = await worker(task_msg)
    return ToolResponse(content=res.content) # Return the worker's response as the tool response

toolkit = Toolkit()
toolkit.register_tool_function(create_worker)

Working with AgentScope

This section provides guidance on how to effectively answer questions about AgentScope or coding with the framework.

Step 1: Clone the Repository First

CRITICAL: Before doing anything else, clone or update the AgentScope repository. The repository contains essential examples and references.

# Clone into this skill directory so that you can refer to it across different sessions
cd /path/to/this/skill/directory
git clone -b main https://github.com/agentscope-ai/agentscope.git

# Or update if already cloned
cd /path/to/this/skill/directory/agentscope
git pull

Why this matters: The repository contains working examples, complete API documentation in source code, and implementation patterns that are more reliable than guessing.

Step 2: Understand the Repository Structure

The cloned repository is organized as follows. Note this may be outdated as the project evolves, you should always check the actual structure after cloning.

agentscope/
├── src/agentscope/          # Main library source code
│   ├── agent/               # Agent implementations (ReActAgent, etc.)
│   ├── model/               # LLM API wrappers (OpenAI, Anthropic, DashScope, etc.)
│   ├── formatter/           # Message formatters for different models
│   ├── memory/              # Memory implementations
│   ├── tool/                # Tool management and built-in tools
│   ├── message/             # Msg class and content blocks
│   ├── pipeline/            # Multi-agent orchestration (MsgHub, etc.)
│   ├── session/             # Session/state management
│   ├── mcp/                 # MCP integration
│   ├── rag/                 # RAG functionality
│   ├── realtime/            # Realtime voice interaction
│   ├── tts/                 # Text-to-speech
│   ├── evaluate/            # Evaluation tools
│   └── ...                  # Other modules
│
├── examples/                # Working examples organized by category
│   ├── agent/               # Different agent types
│   │   └── ...
│   ├── workflows/           # Multi-agent workflows
│   │   └── ...
│   ├── functionality/       # Specific features
│   │   └── ...
│   ├── deployment/          # Deployment patterns
│   ├── integration/         # Third-party integrations
│   ├── evaluation/          # Evaluation examples
│   └── game/                # Game examples (e.g., werewolves)
│
├── docs/                    # Documentation
│   ├── tutorial/            # Tutorial markdown files
│   ├── changelog.md         # Version history
│   └── roadmap.md           # Development roadmap
│
└── tests/                   # Test files

Step 3: Browse Examples by Category

When looking for similar implementations, browse the examples directory by category rather than searching by keywords alone:

1. Start with the category that matches your use case:

- Building a specific agent type? → examples/agent/ - Multi-agent system? → examples/workflows/ - Need a specific feature (MCP, RAG, session)? → examples/functionality/ - Deployment patterns? → examples/deployment/

1. List the subdirectories to see what's available:

- Use file listing tools to explore directory structure - Read directory names to understand what each example covers

1. Read example files to understand implementation patterns:

- Most examples contain a main script and supporting files - Look for README files in subdirectories for explanations

1. Combine with text search when needed:

- After identifying relevant directories, search within them for specific patterns - Search for class names, method calls, or specific functionality

Example workflow:

User asks: "Build a FastAPI app with AgentScope"
→ Browse: List files in examples/deployment/
→ Check: Are there any web service examples?
→ Search: Look for "fastapi", "flask", "api", "server" in examples/
→ Read: Found examples and adapt to user's needs

Step 4: Verify Functionality Exists

Before implementing custom solutions, verify if AgentScope already provides the functionality:

1. List required functionalities (e.g., session management, MCP integration, RAG)
2. Check if provided:

- Browse examples for examples - Search tutorial documentation in docs/tutorial/ - Use the provided scripts (see Part 3) to explore API structure - Read source code in src/agentscope/ for implementation details

1. If not provided: Check how to customize by reading base classes and inheritance patterns in source code

Step 5: Make a Plan

Always create a plan before coding:

1. Identify what AgentScope components you'll use
2. Determine what needs custom implementation
3. Outline the architecture and data flow
4. Consider edge cases and error handling

Step 6: Code with API Reference

When writing code:

1. Check docstrings and arguments before using any class/method

- Read source code files to see signatures and documentation, or - Use the provided scripts to view module/class structures - NEVER make up classes, methods, or arguments

1. Check parent classes - A class's functionality includes inherited methods
2. Manage lifecycle - Clean up resources when needed (close connections, release memory)

Common Pitfalls to Avoid

• ❌ Guessing API signatures without checking documentation
• ❌ Implementing features that already exist in AgentScope
• ❌ Mixing incompatible Model and Formatter (e.g., OpenAI model with DashScope formatter)
• ❌ Forgetting to await async agent calls
• ❌ Not checking parent class methods when searching for functionality
• ❌ Searching by keywords only without browsing the organized examples directory structure

Resources

This section lists all available resources for working with AgentScope.

Official Documentation

• Tutorial: Comprehensive step-by-step guide covering most functionalities in detail. This is the primary resource for learning AgentScope.

GitHub Resources

• Main Repository: Source code, examples, and documentation
• Project Board: Official development roadmap and task tracking
• Design Discussions: In-depth explanations about specific modules/functions/components

Repository Structure

When the repository is cloned locally, the following structure is available for reference:

• src/agentscope/: Main library source code

- Read this for API implementation details - Check docstrings for parameter descriptions - Understand inheritance hierarchies

• examples/: Working examples demonstrating features

- Start here when building similar applications - Examples cover: basic agents, multi-agent systems, tool usage, deployment patterns

• docs/tutorial/: Tutorial documentation source files

- Markdown files explaining concepts and usage - More detailed than README files

Scripts

Located in scripts/ directory of this skill.

• view_pypi_latest_version.sh: View the latest version of AgentScope on PyPI.

cd /path/to/this/skill/directory/scripts/
bash view_pypi_latest_version.sh

• view_module_signature.py: Explore the structure of AgentScope modules, classes, and methods.

Search strategy: Use deep-first search - start broad, then narrow down:

1. agentscope → see all submodules
2. agentscope.agent → see agent-related classes
3. agentscope.agent.ReActAgent → see specific class methods

cd /path/to/this/skill/directory/scripts/
# View top-level module
python view_module_signature.py --module agentscope
# View specific submodule
python view_module_signature.py --module agentscope.agent
# View specific class
python view_module_signature.py --module agentscope.agent.ReActAgent

Reference

Located in references/ directory of this skill.

• multi_agent_orchestration.md: Multi-agent orchestration concepts and implementation
• deployment_guide.md: Deployment patterns and best practices