2nd/10_Wiki/Topics/AI_and_ML/_뇌와 팔다리의 분리_ - 관심사의 분리 (Separation of Concerns).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, inferred_by, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	inferred_by	tech_stack
wiki-2026-0508-뇌와-팔다리의-분리-관심사의-분리-separation-of	Separation of Concerns (관심사의 분리)	10_Wiki/Topics	verified	self	SoC 관심사의 분리 separation of concerns modularity decoupling brain-limbs	none	B	0.85	conceptual	architecture design-principles modularity decoupling srp hexagonal ddd ai-agent-design		2026-05-09	pending	Claude Opus 4.7 (manual cleanup 2026-05-09)	language applicable_to TS / Python / generic Software Architecture AI Agent System Design

Separation of Concerns (관심사의 분리)

📌 한 줄 통찰 (The Karpathy Summary)

"매 module 의 own 역할 — 1 reason to change". Dijkstra 1974. SRP / DDD bounded context / hexagonal / module boundary 의 base. AI agent 의 decision loop + execution tool 의 분리 의 modern 응용.

📖 구조화된 지식 (Synthesized Content)

정의

Edsger Dijkstra (1974) "On the role of scientific thought":

"Separation of concerns... is what I sometimes have called 'the focusing of attention upon some aspect': it does not mean ignoring the other aspects, it is just that from this aspect's point of view, the other is irrelevant."

→ 매 aspect 의 focus + 매 다른 aspect 의 irrelevant.

핵심 idea

1. Modularity: 매 module 의 specific responsibility.
2. Decoupling: 매 change 의 ripple ↓.
3. Boundaries: 매 concern 의 explicit boundary.
4. Maintainability: 매 code 의 understand + modify.

"뇌와 팔다리"

• 뇌 (decision): business logic, strategy.
• 팔다리 (execution): tool, IO, side effect.

→ 매 brain 의 단일 logic. 매 limb 의 swap 가능.

매 응용

MVC (Model-View-Controller)

• Model: data + business rule.
• View: presentation.
• Controller: input + flow.

Hexagonal / Clean Architecture

• Domain: business rule (core).
• Application: use case.
• Infrastructure: DB, API, UI.

→ 매 inner 의 outer 무관. 매 outer 의 swap.

DDD (Domain-Driven Design)

• Bounded Context: 매 domain 의 own model.
• 매 different context 의 own language.
• 매 ACL (Anti-Corruption Layer) 의 boundary.

Microservices

• 매 service 의 own responsibility.
• 매 service 의 own DB.
• 매 service 의 own deploy.

Single Responsibility Principle (SRP)

• 매 class / module 의 1 reason to change.
• 매 specific actor / use case 의 own.

매 game design

Game logic vs rendering

• Game logic: rule, state, behavior.
• Rendering: visual, animation.

→ 매 server (no rendering) + 매 client (rendering only).

AI vs game state

• AI: decision (behavior tree, FSM).
• Game state: data.

→ 매 AI 의 different (player, enemy, NPC) + 매 same data structure.

매 AI agent

Decision loop vs execution

옛 monolithic: 매 logic + execution mixed.

def agent_loop():
    # Logic + tool call mixed
    if some_condition:
        result = direct_db_query(...)
    elif other:
        result = direct_api_call(...)
    return decide(result)

현대 SoC: 매 decision 의 LLM, 매 tool 의 separate.

class Agent:
    def think(self, context):
        # 매 reasoning (no side effect)
        return llm.complete(context)
    
    def execute(self, action):
        # 매 tool call (no reasoning)
        return self.tools[action.name](action.input)
    
    def loop(self):
        while not done:
            decision = self.think(context)
            result = self.execute(decision.action)
            context = update(context, result)

→ 매 brain (LLM) + 매 limb (tool) 의 swap 가능.

매 모범 example

React component

// ❌ Mixed
function UserProfile({ userId }) {
  const [user, setUser] = useState(null);
  
  useEffect(() => {
    fetch(`/api/users/${userId}`)
      .then(r => r.json())
      .then(data => {
        const formatted = `${data.firstName} ${data.lastName.toUpperCase()}`;
        setUser({ ...data, fullName: formatted });
      });
  }, [userId]);
  
  return user ? <h1>{user.fullName}</h1> : <Spinner />;
}

// ✅ Separated
function useUser(userId) {
  return useQuery(['user', userId], () => api.getUser(userId));
}

function formatUser(user) {
  return { ...user, fullName: `${user.firstName} ${user.lastName.toUpperCase()}` };
}

function UserProfile({ userId }) {
  const { data: user, isLoading } = useUser(userId);
  if (isLoading) return <Spinner />;
  return <h1>{formatUser(user).fullName}</h1>;
}

→ 매 fetching (hook) + 매 formatting (function) + 매 rendering (component) 의 분리.

Backend API

# ❌ Mixed
@app.post('/orders')
def create_order(req):
    user = db.execute('SELECT ... FROM users WHERE id = ?', req.user_id).fetchone()
    if user.balance < req.total: return 400
    db.execute('INSERT INTO orders ...')
    db.execute('UPDATE users SET balance = ?', user.balance - req.total)
    send_email(user.email, 'Order confirmed')
    return 200

# ✅ Separated
@app.post('/orders')
def create_order(req):  # HTTP layer
    return order_service.create(req)

class OrderService:  # Business logic
    def __init__(self, user_repo, order_repo, email):
        self.users = user_repo
        self.orders = order_repo
        self.email = email
    
    def create(self, req):
        user = self.users.find(req.user_id)
        if user.balance < req.total: raise InsufficientFundsError()
        
        order = self.orders.create(...)
        self.users.deduct_balance(req.user_id, req.total)
        self.email.send_confirmation(user.email, order)
        return order

class UserRepository:  # Data access
    def find(self, user_id): ...
    def deduct_balance(self, user_id, amount): ...

→ 매 HTTP / business / data 의 layer 별 분리.

매 SoC 의 limit

Over-engineering

• 매 small project 의 9-layer architecture = boilerplate.
• 매 explicit boundary 의 cost.

Coordination cost

• 매 cross-concern feature 의 매 module 의 touch.
• 매 small change 의 5 file.

Pure SoC 의 myth

• 매 real system 의 some coupling.
• 매 perfect 의 X.

→ Pragmatic SoC + 매 case-by-case.

Modern decision

Modular monolith first

• 매 module 의 boundary.
• 매 explicit 1 deploy.
• 매 team 의 ownership.
• 매 future 의 microservice 의 가능.

→ Microservice premature 의 avoid.

매 function (FaaS) extreme

• 매 function 의 own deploy.
• 매 cold start cost.
• 매 distributed 의 complexity.

→ 매 large org 만.

매 monolith with internal SoC

• 매 1 codebase + 매 module / package boundary.
• 매 lint rule (ESLint boundary).
• 매 explicit interface.

→ 매 most common modern.

💻 코드 패턴 (Code Patterns)

Module boundary (TS)

// modules/users/index.ts
export { User } from './domain/User';
export { UserService } from './application/UserService';
// 매 internal 가 not export

// modules/orders/application/OrderService.ts
import { UserService } from '../../users';   // ✅ public
import { internal } from '../../users/domain/secret';   // ❌ ban

ESLint boundary plugin

// .eslintrc.json
{
  "plugins": ["boundaries"],
  "rules": {
    "boundaries/element-types": ["error", {
      "default": "disallow",
      "rules": [
        { "from": "feature/*/ui", "allow": ["feature/*/api"] },
        { "from": "feature/*/api", "allow": ["shared"] }
      ]
    }]
  }
}

Dependency injection (FastAPI)

from fastapi import Depends

def get_user_repo() -> UserRepository:
    return UserRepository(db)

def get_order_service(user_repo = Depends(get_user_repo)) -> OrderService:
    return OrderService(user_repo, ...)

@app.post('/orders')
def create_order(req: CreateOrderReq, service: OrderService = Depends(get_order_service)):
    return service.create(req)

Hexagonal port + adapter (TS)

// Domain (port)
interface PaymentGateway {
  charge(amount: Money, card: CardToken): Promise<Payment>;
}

// Adapter
class StripeGateway implements PaymentGateway {
  async charge(amount, card) { /* Stripe */ }
}

class FakeGateway implements PaymentGateway {
  async charge(amount, card) { /* test */ }
}

// Service
class CheckoutService {
  constructor(private gateway: PaymentGateway) {}
  
  async checkout(...) {
    const payment = await this.gateway.charge(...);
  }
}

// Wiring
const service = new CheckoutService(new StripeGateway());

→ 매 production = Stripe, 매 test = Fake. 매 service 의 unchanged.

AI agent decision / execution

class Agent:
    def __init__(self, llm, tools: dict):
        self.llm = llm
        self.tools = tools
    
    def think(self, context):
        """Pure decision (no side effect)."""
        return self.llm.complete(
            system="You are an agent. Decide next action.",
            user=context,
            tools=list(self.tools.keys()),
        )
    
    def execute(self, action):
        """Pure execution (no decision)."""
        if action.name not in self.tools:
            raise UnknownTool(action.name)
        return self.tools[action.name](action.input)
    
    def loop(self, task, max_steps=10):
        context = task
        for _ in range(max_steps):
            decision = self.think(context)
            if decision.is_done: return decision.answer
            result = self.execute(decision.action)
            context += f"\n{decision.action.name} returned: {result}"
        raise MaxStepsExceeded()

→ 매 brain (LLM swap 가능) + 매 limb (tool swap).

Game logic (Unity ECS-like)

// ❌ Mixed
public class Player : MonoBehaviour {
    int health = 100;
    void Update() {
        // Input
        if (Input.GetKey(KeyCode.W)) transform.position += Vector3.forward * Time.deltaTime;
        
        // Combat
        if (Physics.OverlapSphere(transform.position, 1).Length > 0) health -= 10;
        
        // Render
        GetComponent<MeshRenderer>().material.color = health < 30 ? Color.red : Color.white;
    }
}

// ✅ Separated systems
public class InputSystem { void Update() {} }
public class MovementSystem { void Update() {} }
public class CombatSystem { void Update() {} }
public class RenderSystem { void Update() {} }

// 매 component 의 only data
public struct Health { public int value; }
public struct Position { public Vector3 value; }

→ 매 system 의 own update. 매 component 의 only data.

React feature folder

src/features/
├── auth/
│   ├── api/        # data layer
│   ├── domain/     # business rule
│   ├── ui/         # component
│   └── index.ts    # public API
├── orders/
│   └── ...
└── shared/         # cross-cutting

→ 매 feature 의 own slice. 매 internal 의 own.

Redux slice (state management)

// 매 feature 의 own slice
const authSlice = createSlice({
  name: 'auth',
  initialState,
  reducers: {
    login: (state, action) => { ... },
    logout: (state) => { ... },
  },
});

const userSlice = createSlice({
  name: 'user',
  initialState,
  reducers: { ... },
});

// 매 slice 의 internal. 매 selector 의 cross-slice.

Test의 SoC

// Unit (logic only)
test('formatUser', () => {
  expect(formatUser({ firstName: 'Alice', lastName: 'doe' }))
    .toEqual({ firstName: 'Alice', lastName: 'doe', fullName: 'Alice DOE' });
});

// Integration (with DB)
test('UserService.create', async () => {
  const service = new UserService(testDb);
  const user = await service.create({...});
  expect(user.id).toBeDefined();
});

// E2E (full stack)
test('signup flow', async () => {
  await page.goto('/signup');
  await page.fill('[name=email]', 'a@x');
  await page.click('button[type=submit]');
  await expect(page).toHaveURL('/dashboard');
});

→ 매 layer 의 own test type.

🤔 의사결정 기준 (Decision Criteria)

상황	추천
Small project	Light SoC (folder structure)
Mid-size	Module + lint rule
Large monolith	Modular monolith + DDD
Microservice	Bounded context per service
AI agent	Decision (LLM) + execution (tool)
React app	Feature folder + custom hook
Game	ECS / system separation

기본값: Module boundary + 매 layer (UI / domain / data) 의 separate. 매 case 의 over-engineer X.

⚠️ 모순 및 업데이트 (Contradictions & Updates)

• Pure SoC vs DRY: 매 cross-concern feature 의 매 layer 의 duplicate. 매 trade-off.
• Boundary 의 cost: 매 explicit boundary 의 boilerplate.
• Microservice premature: 매 small team 의 microservice = pain.
• DDD 의 learning curve: 매 small project 의 overkill.
• AI agent 의 emerging: 매 LLM + tool 의 SoC 의 modern.

🔗 지식 연결 (Graph)

• 부모: Software-Architecture · Design-Principles · Modularity
• 변형: Single-Responsibility-Principle · Bounded-Context-DDD · Hexagonal-Architecture · Clean-Architecture
• 응용: MVC · Modular-Monolith · Microservices · Feature-Folder
• AI: AI-Agent-Architecture · Tool-Composition-Deep · Multi-Agent-Coordination
• 비판: Over-Engineering · Premature-Microservice · Coordination-Cost
• Adjacent: Decoupling · Encapsulation · Information-Hiding · Module-Boundaries

🤖 LLM 활용 힌트 (How to Use This Knowledge)

언제 이 지식을 쓰는가:

• 매 architecture 의 review.
• 매 module / service 의 boundary 결정.
• 매 AI agent 의 decision / execution 분리.
• 매 testability 의 design.
• 매 refactor 의 strategy.

언제 쓰면 안 되는가:

• 매 small script (overkill).
• 매 prototype (premature).
• 매 specific micro-optimization.
• Functional language 의 idiom (different paradigm).

❌ 안티패턴 (Anti-Patterns)

• God class / module: 매 모든 거 1 곳.
• Circular dependency: SoC violation.
• Anemic domain model: 매 logic 의 service. 매 model 의 data 만.
• Service locator everywhere: 매 hidden dependency.
• Pure SoC + 매 small project: over-engineer.
• AI agent 의 mixed decision + execution: untestable.
• Microservice premature: 매 single team 의 distributed pain.

🧪 검증 상태 (Validation)

• 정보 상태: verified (concept-level).
• 출처 신뢰도: B (Dijkstra 1974, "Clean Architecture" Martin, "DDD" Evans).
• 검토 이유: Manual cleanup. Foundational concept 의 stable.

🧬 중복 검사 (Duplicate Check)

• 기존 유사 문서: Single-Responsibility-Principle (subset), Hexagonal-Clean (응용), Modular-Monolith (응용).
• 처리 방식: KEEP (foundational principle).
• 처리 이유: 매 specific architecture 의 base.

🕓 변경 이력 (Changelog)

날짜	변경 내용	처리 방식	신뢰도
2026-05-08	P-Reinforce Phase 1 정규화	UPDATE	A
2026-05-09	Manual cleanup — 매 architecture 응용 + AI agent SoC + code pattern + 안티패턴 추가	UPDATE	B