# 다중 에이전트 오케스트레이션 시스템 - Foundry Local

Microsoft Foundry Local이 제공하는 고급 다중 에이전트 시스템으로, 지능형 에이전트 조정, 전문화된 작업 위임, 협업 문제 해결 패턴을 시연합니다.

## 개요

이 샘플은 Foundry Local을 사용하여 정교한 AI 에이전트 시스템을 구축하는 방법을 보여주며, 함수 호출, 에이전트 오케스트레이션, 협업 AI 워크플로우에 대한 Microsoft의 공식 패턴을 구현합니다.

## 아키텍처

```
┌─────────────────────────────────────────────────────────────────┐
│                    Agent Orchestration System                   │
├─────────────────┬─────────────────┬─────────────────┬───────────┤
│  Coordinator    │   Specialist    │    Function     │  Context  │
│     Agent       │     Agents      │     Registry    │  Manager  │
│                 │                 │                 │           │
│ • Task Analysis │ • Code Expert   │ • Tool Calling  │ • Memory  │
│ • Agent Router  │ • Data Analyst  │ • Validation    │ • History │
│ • Workflow Mgmt │ • Research Bot  │ • Error Handle  │ • State   │
│ • Result Merge  │ • Writing Aid   │ • Type Safety   │ • Context │
└─────────────────┴─────────────────┴─────────────────┴───────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                Microsoft Foundry Local Service                  │
│                                                                 │
│ • Multi-Model Support     • Function Calling API               │
│ • Concurrent Inference    • Tool Integration                   │
│ • Context Preservation    • Performance Monitoring             │
└─────────────────────────────────────────────────────────────────┘
```

## 주요 기능

### 🤖 **지능형 에이전트 조정**
- 동적 작업 분석 및 에이전트 선택
- 자동 작업 분배
- 결과 집계 및 종합
- 에이전트 간 통신 프로토콜

### 🔧 **전문화된 에이전트 유형**
- **Code Expert**: 프로그래밍, 디버깅, 코드 리뷰
- **Data Analyst**: 데이터 처리, 시각화, 인사이트 도출
- **Research Assistant**: 정보 수집, 요약
- **Writing Specialist**: 콘텐츠 작성, 편집, 문서화
- **Problem Solver**: 복잡한 추론, 의사 결정

### ⚡ **고급 함수 호출**
- Microsoft Foundry Local 함수 호출 패턴
- 타입 안전한 도구 정의
- 자동 매개변수 검증
- 오류 처리 및 복구
- 도구 체이닝 및 구성

### 🎯 **스마트 작업 라우팅**
- 의도 분류 및 분석
- 에이전트 능력 매칭
- 부하 분산 및 최적화
- 대체 및 중복 처리

## 사전 요구사항

### 시스템 요구사항
- **Python**: asyncio 지원이 포함된 3.9 이상
- **메모리**: 다중 에이전트를 위해 16GB 이상 권장
- **저장소**: 다중 모델을 위해 15GB 이상
- **CPU/GPU**: 멀티코어 프로세서, GPU 권장

### 종속성
```bash
pip install foundry-local-sdk openai aiohttp asyncio pydantic rich typer
```

### Foundry Local 설정
```powershell
# Install and verify Foundry Local
winget install Microsoft.FoundryLocal
foundry --version

# Download recommended models for agents
foundry model download phi-4-mini
foundry model download qwen2.5-coder-0.5b
foundry model download phi-3.5-mini
```

## 빠른 시작

### 1. 기본 다중 에이전트 워크플로우
```python
from agentic_system import AgentOrchestrator, CodeAgent, ResearchAgent

# Initialize the orchestrator
orchestrator = AgentOrchestrator()

# Add specialized agents
await orchestrator.add_agent(CodeAgent("phi-4-mini"))
await orchestrator.add_agent(ResearchAgent("qwen2.5-coder-0.5b"))

# Execute a complex task
result = await orchestrator.execute_task(
    "Create a Python script that analyzes web traffic data and generates a report"
)

print(result.summary)
```

### 2. 사용자 정의 에이전트 생성
```python
from agentic_system import BaseAgent, tool

class DataAnalystAgent(BaseAgent):
    """Specialized agent for data analysis tasks."""
    
    @tool
    async def analyze_dataset(self, data_path: str, analysis_type: str) -> dict:
        """Analyze a dataset and return insights."""
        # Implementation here
        pass
    
    @tool
    async def create_visualization(self, data: dict, chart_type: str) -> str:
        """Create data visualizations."""
        # Implementation here
        pass

# Use the custom agent
agent = DataAnalystAgent("qwen2.5-0.5b")
result = await agent.analyze_dataset("sales_data.csv", "trend_analysis")
```

### 3. 함수 호출 통합
```python
# Define tools following Microsoft patterns
tools = [
    {
        "name": "web_search",
        "description": "Search the web for information",
        "parameters": {
            "query": {"description": "Search query", "type": "string"},
            "max_results": {"description": "Maximum results", "type": "integer"}
        }
    },
    {
        "name": "code_analyzer", 
        "description": "Analyze code quality and suggest improvements",
        "parameters": {
            "code": {"description": "Code to analyze", "type": "string"},
            "language": {"description": "Programming language", "type": "string"}
        }
    }
]

# Register tools with orchestrator
orchestrator.register_tools(tools)
```

## 프로젝트 구조

```
09/
├── README.md                    # This documentation
├── requirements.txt             # Python dependencies  
├── agentic_system/
│   ├── __init__.py             # Package initialization
│   ├── orchestrator.py         # Main orchestrator class
│   ├── base_agent.py           # Base agent implementation
│   ├── specialized_agents/
│   │   ├── __init__.py
│   │   ├── code_agent.py       # Programming specialist
│   │   ├── research_agent.py   # Research specialist  
│   │   ├── data_agent.py       # Data analysis specialist
│   │   ├── writing_agent.py    # Content creation specialist
│   │   └── solver_agent.py     # Problem solving specialist
│   ├── tools/
│   │   ├── __init__.py
│   │   ├── function_registry.py # Tool management
│   │   ├── web_tools.py        # Web interaction tools
│   │   ├── file_tools.py       # File system tools
│   │   ├── code_tools.py       # Code analysis tools
│   │   └── data_tools.py       # Data processing tools
│   ├── coordination/
│   │   ├── __init__.py
│   │   ├── task_router.py      # Task routing logic
│   │   ├── result_merger.py    # Result aggregation
│   │   ├── context_manager.py  # Context and memory
│   │   └── workflow_engine.py  # Workflow management
│   └── utils/
│       ├── __init__.py
│       ├── foundry_client.py   # Foundry Local integration
│       ├── logging_config.py   # Logging setup
│       └── validation.py       # Input validation
├── examples/
│   ├── basic_coordination.py   # Simple multi-agent example
│   ├── complex_workflow.py     # Advanced workflow example
│   ├── custom_agents.py        # Custom agent creation
│   ├── function_calling.py     # Tool integration example
│   └── interactive_demo.py     # Interactive demonstration
├── tools/
│   ├── web_search.py          # Web search implementation
│   ├── code_analyzer.py       # Code analysis tools
│   ├── data_processor.py      # Data processing tools
│   └── file_manager.py        # File system operations
└── tests/
    ├── test_orchestrator.py   # Orchestrator tests
    ├── test_agents.py         # Agent tests
    ├── test_tools.py          # Tool tests
    └── test_integration.py    # Integration tests
```

## 에이전트 유형 심층 분석

### 1. Code Expert 에이전트
```python
class CodeAgent(BaseAgent):
    """Expert in programming, debugging, and code review."""
    
    specialties = [
        "code_generation", "debugging", "code_review", 
        "refactoring", "testing", "documentation"
    ]
    
    @tool
    async def generate_code(self, specification: str, language: str) -> str:
        """Generate code from specifications."""
        
    @tool  
    async def debug_code(self, code: str, error_message: str) -> dict:
        """Debug code and suggest fixes."""
        
    @tool
    async def review_code(self, code: str, criteria: list) -> dict:
        """Perform comprehensive code review."""
```

### 2. Research Assistant 에이전트
```python
class ResearchAgent(BaseAgent):
    """Specialized in information gathering and analysis."""
    
    specialties = [
        "web_research", "information_synthesis", "fact_checking",
        "summarization", "trend_analysis"
    ]
    
    @tool
    async def research_topic(self, topic: str, depth: str) -> dict:
        """Research a topic comprehensively."""
        
    @tool
    async def summarize_information(self, sources: list, style: str) -> str:
        """Summarize information from multiple sources."""
        
    @tool
    async def fact_check(self, claims: list) -> dict:
        """Verify factual claims."""
```

### 3. Data Analysis 에이전트
```python
class DataAgent(BaseAgent):
    """Expert in data processing and analysis."""
    
    specialties = [
        "data_analysis", "statistical_analysis", "visualization",
        "pattern_recognition", "predictive_modeling"
    ]
    
    @tool
    async def analyze_data(self, dataset: str, analysis_type: str) -> dict:
        """Perform data analysis."""
        
    @tool
    async def create_visualization(self, data: dict, viz_type: str) -> str:
        """Create data visualizations."""
        
    @tool
    async def statistical_test(self, data: dict, test_type: str) -> dict:
        """Perform statistical tests."""
```

## 오케스트레이션 패턴

### 1. 순차 워크플로우
```python
# Define a sequential workflow
workflow = orchestrator.create_workflow("sequential")
workflow.add_step("research", ResearchAgent, "gather_requirements")
workflow.add_step("design", CodeAgent, "create_architecture")  
workflow.add_step("implement", CodeAgent, "write_code")
workflow.add_step("test", CodeAgent, "create_tests")

result = await workflow.execute("Build a REST API for user management")
```

### 2. 병렬 실행
```python
# Execute tasks in parallel
parallel_tasks = [
    ("research_market", ResearchAgent, "analyze_market_trends"),
    ("analyze_competitors", DataAgent, "competitor_analysis"),
    ("technical_feasibility", CodeAgent, "assess_technical_requirements")
]

results = await orchestrator.execute_parallel(parallel_tasks)
synthesized = await orchestrator.synthesize_results(results)
```

### 3. 동적 에이전트 선택
```python
# Automatic agent selection based on task analysis
task = "Create a machine learning model to predict customer churn"

# Orchestrator analyzes task and selects appropriate agents
selected_agents = await orchestrator.analyze_task_requirements(task)
# Returns: [DataAgent, CodeAgent, ResearchAgent]

result = await orchestrator.execute_with_agents(task, selected_agents)
```

## 함수 호출 통합

### Microsoft Foundry Local 패턴
```python
# Define tools following Microsoft's function calling schema
def define_foundry_tools():
    return [
        {
            "name": "analyze_code_quality",
            "description": "Analyze code quality and suggest improvements",
            "parameters": {
                "code": {
                    "description": "The source code to analyze",
                    "type": "string"
                },
                "language": {
                    "description": "Programming language",
                    "type": "string"
                },
                "criteria": {
                    "description": "Analysis criteria",
                    "type": "array",
                    "items": {"type": "string"}
                }
            }
        },
        {
            "name": "search_documentation",
            "description": "Search technical documentation",
            "parameters": {
                "query": {"description": "Search query", "type": "string"},
                "source": {"description": "Documentation source", "type": "string"}
            }
        }
    ]

# Integration with Foundry Local
async def setup_function_calling():
    tools = define_foundry_tools()
    
    # Configure Foundry Local for function calling
    client = openai.OpenAI(
        base_url=manager.endpoint,
        api_key=manager.api_key
    )
    
    # Use tools in conversation
    response = await client.chat.completions.create(
        model=manager.get_model_info("phi-4-mini").id,
        messages=[
            {"role": "user", "content": "Analyze this Python code for quality issues"}
        ],
        tools=[{"type": "function", "function": tool} for tool in tools],
        tool_choice="auto"
    )
```

## 고급 조정 기능

### 1. 컨텍스트 관리
```python
class ContextManager:
    """Manages shared context across agents."""
    
    async def share_context(self, agent_id: str, context: dict):
        """Share context with specific agent."""
        
    async def get_shared_memory(self) -> dict:
        """Retrieve shared memory state."""
        
    async def update_global_state(self, updates: dict):
        """Update global orchestrator state."""
```

### 2. 결과 종합
```python
class ResultMerger:
    """Intelligently merge results from multiple agents."""
    
    async def merge_analyses(self, results: list) -> dict:
        """Merge analysis results."""
        
    async def resolve_conflicts(self, conflicting_results: list) -> dict:
        """Resolve conflicting agent outputs."""
        
    async def create_summary(self, all_results: dict) -> str:
        """Create comprehensive summary."""
```

### 3. 품질 보증
```python
class QualityController:
    """Ensures output quality and consistency."""
    
    async def validate_output(self, result: dict, criteria: list) -> bool:
        """Validate agent output quality."""
        
    async def cross_check_facts(self, claims: list) -> dict:
        """Cross-verify facts across agents."""
        
    async def ensure_consistency(self, outputs: list) -> dict:
        """Ensure consistent outputs."""
```

## 성능 최적화

### 1. 모델 부하 분산
```python
# Distribute models across agents for optimal resource usage
model_allocation = {
    "code_tasks": "phi-4-mini",
    "research_tasks": "qwen2.5-coder-0.5b", 
    "analysis_tasks": "phi-3.5-mini",
    "general_tasks": "phi-4-mini"
}

orchestrator.configure_model_allocation(model_allocation)
```

### 2. 캐싱 및 메모리
```python
# Implement intelligent caching
cache_config = {
    "response_cache": True,
    "context_cache": True,
    "tool_result_cache": True,
    "cache_ttl": 3600  # 1 hour
}

orchestrator.configure_caching(cache_config)
```

### 3. 동시 실행
```python
# Optimize for parallel processing
concurrency_config = {
    "max_concurrent_agents": 4,
    "agent_pool_size": 8,
    "task_queue_size": 100,
    "timeout_seconds": 300
}

orchestrator.configure_concurrency(concurrency_config)
```

## 사용 예제

### 예제 1: 소프트웨어 개발 워크플로우
```python
async def software_development_workflow():
    """Complete software development using multiple agents."""
    
    # Initialize orchestrator with specialized agents
    orchestrator = AgentOrchestrator()
    await orchestrator.add_agent(ResearchAgent("qwen2.5-coder-0.5b"))
    await orchestrator.add_agent(CodeAgent("phi-4-mini"))
    await orchestrator.add_agent(DataAgent("phi-3.5-mini"))
    
    # Define the development task
    task = """
    Create a web application that:
    1. Analyzes user behavior data
    2. Provides real-time analytics dashboard
    3. Includes user authentication
    4. Has comprehensive tests
    """
    
    # Execute coordinated workflow
    result = await orchestrator.execute_workflow(
        task=task,
        workflow_type="software_development",
        quality_gates=["code_review", "testing", "security_check"]
    )
    
    return result
```

### 예제 2: 연구 및 분석
```python
async def comprehensive_research():
    """Multi-agent research coordination."""
    
    research_query = "Impact of AI on software development productivity"
    
    # Parallel research execution
    tasks = [
        ("literature_review", ResearchAgent, research_query),
        ("data_analysis", DataAgent, "productivity_metrics"),
        ("case_studies", ResearchAgent, "ai_adoption_cases"),
        ("technical_analysis", CodeAgent, "ai_tool_evaluation")
    ]
    
    results = await orchestrator.execute_parallel(tasks)
    
    # Synthesize findings
    final_report = await orchestrator.synthesize_research(
        results=results,
        format="comprehensive_report",
        include_recommendations=True
    )
    
    return final_report
```

### 예제 3: 문제 해결 세션
```python
async def collaborative_problem_solving():
    """Multi-agent collaborative problem solving."""
    
    problem = """
    A company's API response times have increased 300% over the past month.
    Analyze the issue and propose solutions.
    """
    
    # Deploy specialist agents
    investigation_plan = await orchestrator.create_investigation_plan(problem)
    
    agents_deployed = [
        (CodeAgent, "analyze_code_performance"),
        (DataAgent, "analyze_performance_metrics"), 
        (ResearchAgent, "research_similar_issues"),
        (SolverAgent, "propose_solutions")
    ]
    
    # Coordinate investigation
    findings = await orchestrator.coordinate_investigation(
        problem=problem,
        agents=agents_deployed,
        investigation_plan=investigation_plan
    )
    
    # Generate action plan
    action_plan = await orchestrator.create_action_plan(findings)
    
    return action_plan
```

## 구성 및 사용자 정의

### 에이전트 구성
```python
# Configure individual agents
agent_configs = {
    "CodeAgent": {
        "model": "phi-4-mini",
        "temperature": 0.3,
        "max_tokens": 2000,
        "specialization_level": "expert"
    },
    "ResearchAgent": {
        "model": "qwen2.5-coder-0.5b",
        "temperature": 0.7,
        "max_tokens": 1500,
        "research_depth": "comprehensive"
    }
}

orchestrator.configure_agents(agent_configs)
```

### 워크플로우 사용자 정의
```python
# Custom workflow definitions
custom_workflows = {
    "data_science_project": [
        "data_collection",
        "exploratory_analysis", 
        "model_development",
        "validation_testing",
        "deployment_preparation"
    ],
    "security_audit": [
        "vulnerability_scan",
        "code_review", 
        "penetration_testing",
        "compliance_check",
        "remediation_plan"
    ]
}

orchestrator.register_workflows(custom_workflows)
```

## 모니터링 및 분석

### 성능 추적
```python
# Monitor orchestrator performance
metrics = await orchestrator.get_performance_metrics()
print(f"Tasks Completed: {metrics.tasks_completed}")
print(f"Average Response Time: {metrics.avg_response_time}s")
print(f"Success Rate: {metrics.success_rate}%")
print(f"Agent Utilization: {metrics.agent_utilization}")
```

### 품질 지표
```python
# Track output quality
quality_report = await orchestrator.generate_quality_report()
print(f"Output Consistency: {quality_report.consistency_score}")
print(f"Factual Accuracy: {quality_report.accuracy_score}")
print(f"Completeness: {quality_report.completeness_score}")
```

## 학습 결과

이 샘플을 완료한 후, 다음을 이해할 수 있습니다:

1. **다중 에이전트 시스템 아키텍처**
   - 에이전트 조정 패턴
   - 작업 분배 전략
   - 결과 종합 기술
   - 에이전트 간 컨텍스트 관리

2. **Microsoft Foundry Local 통합**
   - 함수 호출 구현
   - 도구 통합 패턴
   - 다중 모델 오케스트레이션
   - 성능 최적화

3. **고급 AI 오케스트레이션**
   - 워크플로우 설계 및 실행
   - 품질 보증 메커니즘
   - 오류 처리 및 복구
   - 확장성 고려사항

4. **프로덕션 시스템 설계**
   - 모니터링 및 분석
   - 구성 관리
   - 보안 모범 사례
   - 성능 조정

## 다음 단계

- **샘플 10**: Foundry Local을 도구로 통합
- **고급 주제**: 사용자 정의 에이전트 개발
- **확장**: 분산 에이전트 시스템
- **통합**: 엔터프라이즈 워크플로우 통합

## 기여

기여 지침은 메인 저장소 가이드를 참조하세요.

## 라이선스

이 샘플은 Microsoft Foundry Local 프로젝트와 동일한 라이선스를 따릅니다.

---