multi-agent on Astroicers Blog

AutoGen 多 Agent 程式框架

Mon, 25 May 2026 00:00:00 +0000

專案簡介

AutoGen 是 Microsoft 開發的多 Agent 對話框架，讓多個 AI Agent 透過對話協作完成任務。支援程式碼執行、人機協作、自訂 Agent 等功能。

GitHub Stars: 54K+

主要功能

對話式協作 - Agent 透過對話完成任務
程式執行 - 自動執行生成的程式碼
人機協作 - 人類參與 Agent 對話
可擴展 - 自訂 Agent 行為
多模型 - 支援各種 LLM

安裝

1

pip install autogen-agentchat

基本使用

雙 Agent 對話

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28


from autogen import AssistantAgent, UserProxyAgent

# 設定 LLM
config_list = [
 {
 "model": "gpt-4o",
 "api_key": "your-api-key"
 }
]

# 建立 Assistant
assistant = AssistantAgent(
 name="assistant",
 llm_config={"config_list": config_list}
)

# 建立 User Proxy（可執行程式碼）
user_proxy = UserProxyAgent(
 name="user_proxy",
 human_input_mode="NEVER",
 code_execution_config={"work_dir": "coding"}
)

# 開始對話
user_proxy.initiate_chat(
 assistant,
 message="Write a Python script to scan for open ports on localhost"
)

Agent 類型

AssistantAgent

AI 助手，生成回應：

1
2
3
4
5
6


assistant = AssistantAgent(
 name="security_expert",
 system_message="""You are a cybersecurity expert.
 Help with security analysis and provide recommendations.""",
 llm_config={"config_list": config_list}
)

UserProxyAgent

代表使用者，可執行程式碼：

1
2
3
4
5
6
7
8
9


user_proxy = UserProxyAgent(
 name="user",
 human_input_mode="TERMINATE", # 結束時請求輸入
 max_consecutive_auto_reply=10,
 code_execution_config={
 "work_dir": "workspace",
 "use_docker": True
 }
)

GroupChat

多 Agent 群組對話：

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21


from autogen import GroupChat, GroupChatManager

researcher = AssistantAgent(name="researcher", ...)
analyst = AssistantAgent(name="analyst", ...)
writer = AssistantAgent(name="writer", ...)

groupchat = GroupChat(
 agents=[user_proxy, researcher, analyst, writer],
 messages=[],
 max_round=20
)

manager = GroupChatManager(
 groupchat=groupchat,
 llm_config={"config_list": config_list}
)

user_proxy.initiate_chat(
 manager,
 message="Research and write a report on SQL injection attacks"
)

程式碼執行

本地執行

1
2
3
4
5
6
7


user_proxy = UserProxyAgent(
 name="executor",
 code_execution_config={
 "work_dir": "coding",
 "use_docker": False
 }
)

Docker 執行

1
2
3
4
5
6
7


user_proxy = UserProxyAgent(
 name="executor",
 code_execution_config={
 "work_dir": "coding",
 "use_docker": "python:3.11"
 }
)

安全研究範例

漏洞分析團隊

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58


from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager

config_list = [{"model": "gpt-4o", "api_key": "your-key"}]

# 研究員
researcher = AssistantAgent(
 name="Vulnerability_Researcher",
 system_message="""You are a vulnerability researcher.
 Search and analyze security vulnerabilities.
 Provide CVE details and technical analysis.""",
 llm_config={"config_list": config_list}
)

# 分析師
analyst = AssistantAgent(
 name="Risk_Analyst",
 system_message="""You are a risk analyst.
 Assess the impact and exploitability of vulnerabilities.
 Provide risk scores and prioritization.""",
 llm_config={"config_list": config_list}
)

# 修復建議
remediation = AssistantAgent(
 name="Remediation_Advisor",
 system_message="""You are a remediation advisor.
 Provide patches, workarounds, and mitigation strategies.""",
 llm_config={"config_list": config_list}
)

# 執行者
executor = UserProxyAgent(
 name="Executor",
 human_input_mode="NEVER",
 code_execution_config={"work_dir": "security_analysis"}
)

# 群組對話
groupchat = GroupChat(
 agents=[executor, researcher, analyst, remediation],
 messages=[],
 max_round=15
)

manager = GroupChatManager(
 groupchat=groupchat,
 llm_config={"config_list": config_list}
)

# 開始分析
executor.initiate_chat(
 manager,
 message="""Analyze CVE-2026-12345:
 1. Research the vulnerability details
 2. Assess the risk level
 3. Provide remediation recommendations
 """
)

人機協作

請求人類輸入

1
2
3
4
5


user_proxy = UserProxyAgent(
 name="user",
 human_input_mode="ALWAYS", # 每次都請求輸入
 max_consecutive_auto_reply=5
)

選項

ALWAYS - 每次回覆前請求輸入
TERMINATE - 終止對話時請求
NEVER - 完全自動

自訂 Agent

函數呼叫

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13


from autogen import register_function

def scan_ports(target: str, ports: str) -> str:
 """Scan ports on target host"""
 # 實作掃描邏輯
 return f"Scan results for {target}"

register_function(
 scan_ports,
 caller=assistant,
 executor=user_proxy,
 description="Scan network ports on a target"
)

自訂 Agent 類別

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16


from autogen import ConversableAgent

class SecurityAgent(ConversableAgent):
 def __init__(self, name, **kwargs):
 super().__init__(name, **kwargs)
 self.register_reply(
 trigger=self._should_scan,
 reply_func=self._scan_reply
 )

 def _should_scan(self, messages):
 return "scan" in messages[-1]["content"].lower()

 def _scan_reply(self, messages, sender, config):
 # 自訂回覆邏輯
 return True, "Scanning..."

模型設定

多模型

 1
 2
 3
 4
 5
 6
 7
 8
 9
10


config_list = [
 {"model": "gpt-4o", "api_key": "key1"},
 {"model": "gpt-4o-mini", "api_key": "key2"},
]

# 自動 fallback
llm_config = {
 "config_list": config_list,
 "temperature": 0.7
}

本地模型

1
2
3
4
5
6
7


config_list = [
 {
 "model": "llama3.3",
 "base_url": "http://localhost:11434/v1",
 "api_key": "ollama"
 }
]

對話控制

終止條件

1
2
3
4
5
6
7


def is_termination_msg(msg):
 return "TERMINATE" in msg.get("content", "")

assistant = AssistantAgent(
 name="assistant",
 is_termination_msg=is_termination_msg
)

最大回覆數

1
2
3
4


user_proxy = UserProxyAgent(
 name="user",
 max_consecutive_auto_reply=10
)

CrewAI 多 Agent 協作框架

Thu, 21 May 2026 00:00:00 +0000

專案簡介

CrewAI 是一個多 Agent 協作框架，讓你建立由多個專業 AI Agent 組成的團隊。每個 Agent 有不同的角色和技能，透過協作完成複雜任務。

GitHub Stars: 43K+

主要功能

角色扮演 - 定義專業 Agent 角色
任務分配 - 自動分解和分配任務
協作機制 - Agent 間溝通協作
工具整合 - 每個 Agent 可使用不同工具
流程控制 - 順序、並行、層級執行

安裝

1

pip install crewai crewai-tools

基本概念

Agent

Agent 是具有特定角色的 AI 實體：

1
2
3
4
5
6
7
8


from crewai import Agent

researcher = Agent(
 role="Security Researcher",
 goal="Find and analyze security vulnerabilities",
 backstory="Expert in penetration testing with 10 years experience",
 verbose=True
)

Task

Task 定義 Agent 需要完成的工作：

1
2
3
4
5
6
7


from crewai import Task

research_task = Task(
 description="Research the latest CVE vulnerabilities in web applications",
 expected_output="A detailed report of top 5 critical CVEs",
 agent=researcher
)

Crew

Crew 是 Agent 團隊：

1
2
3
4
5
6
7
8
9


from crewai import Crew

crew = Crew(
 agents=[researcher, analyst, writer],
 tasks=[research_task, analysis_task, report_task],
 verbose=True
)

result = crew.kickoff()

完整範例

安全研究團隊

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68


from crewai import Agent, Task, Crew, Process
from crewai_tools import SerperDevTool, WebsiteSearchTool

# 工具
search_tool = SerperDevTool()
web_tool = WebsiteSearchTool()

# Agent 定義
researcher = Agent(
 role="Vulnerability Researcher",
 goal="Find the latest critical security vulnerabilities",
 backstory="""You are a senior security researcher
 specializing in web application vulnerabilities.""",
 tools=[search_tool, web_tool],
 verbose=True
)

analyst = Agent(
 role="Security Analyst",
 goal="Analyze vulnerabilities and assess their impact",
 backstory="""You are a security analyst with expertise
 in risk assessment and threat modeling.""",
 verbose=True
)

writer = Agent(
 role="Technical Writer",
 goal="Create clear security advisories",
 backstory="""You are a technical writer who creates
 actionable security documentation.""",
 verbose=True
)

# Task 定義
research_task = Task(
 description="""Search for the top 5 critical CVEs
 announced in the past month. Focus on web applications.""",
 expected_output="List of CVEs with descriptions",
 agent=researcher
)

analysis_task = Task(
 description="""Analyze the CVEs from the research
 and assess their risk level and impact.""",
 expected_output="Risk assessment for each CVE",
 agent=analyst,
 context=[research_task]
)

report_task = Task(
 description="""Create a security advisory report
 based on the analysis.""",
 expected_output="Markdown formatted security report",
 agent=writer,
 context=[analysis_task]
)

# 建立 Crew
security_crew = Crew(
 agents=[researcher, analyst, writer],
 tasks=[research_task, analysis_task, report_task],
 process=Process.sequential,
 verbose=True
)

# 執行
result = security_crew.kickoff()
print(result)

執行流程

順序執行

1
2
3
4
5


crew = Crew(
 agents=[agent1, agent2, agent3],
 tasks=[task1, task2, task3],
 process=Process.sequential # 依序執行
)

層級執行

1
2
3
4
5
6


crew = Crew(
 agents=[manager, worker1, worker2],
 tasks=[task1, task2, task3],
 process=Process.hierarchical, # 管理者分配任務
 manager_llm=ChatOpenAI(model="gpt-4o")
)

工具整合

內建工具

1
2
3
4
5
6
7


from crewai_tools import (
 SerperDevTool, # 搜尋
 WebsiteSearchTool, # 網站搜尋
 FileReadTool, # 讀取檔案
 DirectoryReadTool, # 讀取目錄
 CodeInterpreterTool # 執行程式
)

自訂工具

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21


from crewai_tools import BaseTool
from pydantic import BaseModel, Field

class ScanInput(BaseModel):
 target: str = Field(description="Target URL to scan")

class VulnerabilityScanner(BaseTool):
 name: str = "Vulnerability Scanner"
 description: str = "Scans a target for security vulnerabilities"
 args_schema: type[BaseModel] = ScanInput

 def _run(self, target: str) -> str:
 # 實作掃描邏輯
 return f"Scan results for {target}"

# 使用
scanner = VulnerabilityScanner()
agent = Agent(
 role="Security Scanner",
 tools=[scanner]
)

進階功能

任務依賴

1
2
3
4
5


task2 = Task(
 description="Analyze the research results",
 agent=analyst,
 context=[task1] # 依賴 task1 的結果
)

輸出格式

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14


from pydantic import BaseModel

class VulnerabilityReport(BaseModel):
 cve_id: str
 severity: str
 description: str
 mitigation: str

task = Task(
 description="Create vulnerability report",
 expected_output="Structured vulnerability report",
 output_pydantic=VulnerabilityReport,
 agent=writer
)

回呼函數

1
2
3
4
5
6
7
8


def on_task_complete(task_output):
 print(f"Task completed: {task_output}")

task = Task(
 description="Research vulnerabilities",
 callback=on_task_complete,
 agent=researcher
)

記憶功能

啟用記憶

1
2
3
4
5
6
7
8
9


crew = Crew(
 agents=[...],
 tasks=[...],
 memory=True, # 啟用記憶
 embedder={
 "provider": "openai",
 "config": {"model": "text-embedding-3-small"}
 }
)

記憶類型

短期記憶 - 當前對話上下文
長期記憶 - 跨會話知識
實體記憶 - 識別的實體和關係

模型設定

使用不同模型

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13


from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

agent = Agent(
 role="Researcher",
 llm=ChatOpenAI(model="gpt-4o")
)

# 或使用 Anthropic
agent = Agent(
 role="Analyst",
 llm=ChatAnthropic(model="claude-3-5-sonnet-20241022")
)

本地模型

1
2
3
4
5
6


from langchain_community.llms import Ollama

agent = Agent(
 role="Researcher",
 llm=Ollama(model="llama3.3")
)

最佳實踐

明確的角色定義

1
2
3
4
5


agent = Agent(
 role="Security Penetration Tester", # 具體角色
 goal="Identify and exploit vulnerabilities", # 明確目標
 backstory="10 years of experience in red team operations" # 背景故事
)

任務粒度

1
2
3
4
5


# ✓ 好的任務定義
Task(description="Find SQL injection vulnerabilities in login form")

# ✗ 太模糊
Task(description="Test security")

multi-agent on Astroicers Blog

AutoGen 多 Agent 程式框架

專案簡介

主要功能

安裝

基本使用

雙 Agent 對話

Agent 類型

AssistantAgent

UserProxyAgent

GroupChat

程式碼執行

本地執行

Docker 執行

安全研究範例

漏洞分析團隊

人機協作

請求人類輸入

選項

自訂 Agent

函數呼叫

自訂 Agent 類別

模型設定

多模型

本地模型

對話控制

終止條件

最大回覆數

相關連結

延伸閱讀

CrewAI 多 Agent 協作框架

專案簡介

主要功能

安裝

基本概念

Agent

Task

Crew

完整範例

安全研究團隊

執行流程

順序執行

層級執行

工具整合

內建工具

自訂工具

進階功能

任務依賴

輸出格式

回呼函數

記憶功能

啟用記憶

記憶類型

模型設定

使用不同模型

本地模型

最佳實踐

明確的角色定義

任務粒度

相關連結

延伸閱讀