1
2
3
4
5
6

graph TD
    A[Task Input] --> B[LLM Reasoning]
    B --> C[Action Execution]
    C --> D[Observation]
    D --> B
    B --> E[Final Output]

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

graph TD
    A[Task] --> B[Task Agent]
    B --> C[Task Result]
    C --> D[Evaluator]
    D --> E[Feedback]
    E --> F[Meta Agent]
    F --> G[Updated Meta Code]
    G --> B
    F --> H[Updated Improvement Mechanism]
    H --> F

 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
69
70
71
72

import json
from typing import Dict, Any
from dataclasses import dataclass, field

@dataclass
class TaskResult:
    """작업 결과를 저장하는 데이터 클래스"""
    task_id: str
    output: str
    score: float
    feedback: str

@dataclass
class MetaCode:
    """메타 수준의 코드 (프롬프트, 전략 등)"""
    version: int = 1
    prompt_template: str = "Solve the following problem: {task}"
    strategy: str = "step_by_step"
    evaluation_criteria: Dict[str, float] = field(
        default_factory=lambda: {"accuracy": 1.0}
    )

class HyperAgent:
    """자기참조적 개선이 가능한 에이전트"""
    
    def __init__(self, llm_client):
        self.llm = llm_client
        self.meta_code = MetaCode()
        self.history = []
    
    def execute_task(self, task: str) -> TaskResult:
        """Task Agent: 메타 코드를 사용하여 작업 수행"""
        prompt = self.meta_code.prompt_template.format(task=task)
        
        # LLM으로 작업 수행
        response = self.llm.generate(prompt)
        
        # 결과 평가
        result = TaskResult(
            task_id=f"task_{len(self.history)}",
            output=response,
            score=0.0,  # 실제로는 Evaluator가 계산
            feedback=""
        )
        
        self.history.append(result)
        return result
    
    def meta_improve(self) -> MetaCode:
        """Meta Agent: 자기 개선 메커니즘 수행"""
        
        # 이전 성능 분석
        performance_summary = self._analyze_performance()
        
        # 메타 프롬프트: 개선 방안 생성
        meta_prompt = f"""
        Current meta code version: {self.meta_code.version}
        Current prompt template: {self.meta_code.prompt_template}
        Current strategy: {self.meta_code.strategy}
        
        Recent performance: {performance_summary}
        
        Generate improved meta code as JSON:
        {{
            "prompt_template": "...",
            "strategy": "...",
            "rationale": "..."
        }}
        """
        
        response = self.llm.generate(meta_prompt)
        improvements = json.

 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

loads(response)
        
        # 메타 코드 업데이트
        new_meta_code = MetaCode(
            version=self.meta_code.version + 1,
            prompt_template=improvements["prompt_template"],
            strategy=improvements["strategy"],
            evaluation_criteria=self.meta_code.evaluation_criteria
        )
        
        # 안전성 검사 후 적용
        if self._validate_meta_code(new_meta_code):
            self.meta_code = new_meta_code
            print(f"Meta code updated to v{self.meta_code.version}")
            print(f"Rationale: {improvements.get('rationale', 'N/A')}")
        
        return self.meta_code
    
    def _analyze_performance(self) -> Dict[str, Any]:
        """성능 이력 분석"""
        if not self.history:
            return {"status": "no_data"}
        
        recent = self.history[-5:]  # 최근 5개 작업
        avg_score = sum(r.score for r in recent) / len(recent)
        
        return {
            "avg_score": avg_score,
            "num_tasks": len(self.history),
            "trend": "improving" if len(self.history) > 1 else "initial"
        }
    
    def _validate_meta_code(self, meta_code: MetaCode) -> bool:
        """메타 코드 안전성 검증"""
        # 프롬프트 길이 제한
        if len(meta_code.prompt_template) > 2000:
            return False
        # 필수 플레이스홀더 확인
        if "{task}" not in meta_code.prompt_template:
            return False
        return True

# 사용 예시
if __name__ == "__main__":
    # 가상의 LLM 클라이언트 (실제로는 OpenAI, Anthropic 등)
    class MockLLM:
        def generate(self, prompt: str) -> str:
            if "improved meta code" in prompt.lower():
                return json.dumps({
                    "prompt_template": (
                        "Analyze and solve step-by-step: {task}
"
                        "Show your reasoning process."
                    ),
                    "strategy": "chain_of_thought",
                    "rationale":

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

 "Adding explicit reasoning improves accuracy"
                })
            return "Sample solution"
    
    agent = HyperAgent(llm_client=MockLLM())
    
    # 작업 수행 → 메타 개선 반복
    for iteration in range(3):
        print(f"
=== Iteration {iteration + 1} ===")
        result = agent.execute_task("Solve: 2x + 5 = 13")
        agent.meta_improve()
        print(f"Current version: v{agent.meta_code.version}")

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

# 초기 메타 코드 정의
initial_meta = {
    "task_type": "code_review",
    "prompt_template": """
    Review the following code for:
    1. Bugs and errors
    2. Performance issues
    3. Best practices
    
    Code: {task}
    """,
    "evaluation_criteria": {
        "bug_detection": 0.4,
        "style_compliance": 0.3,
        "performance_insight": 0.3
    }
}