Function Calling을 넘어서: Tool-Environment 신뢰성 문제 하에서의 Tool-Using Agent 벤치마크

# ToolBench-X 스타일의 장애 진단 + 회복 에이전트 패턴 예시

def tool_call_with_recovery(tool_fn, args, max_retries=2):
    """
    툴 호출 시 장애 유형을 진단하고 적절한 회복 전략을 선택.
    """
    for attempt in range(max_retries + 1):
        try:
            result = tool_fn(**args)
            
            # Output Drift 감지: 예상 필드가 없거나 형식이 다름
            if not validate_output_schema(result):
                result = normalize_output(result)  # 정규화 시도
            
            if result.get('success') is False or result.get('ok') is False:
                raise ValueError(f"Tool returned failure: {result}")
            
            return result
        
        except TimeoutError:
            # Execution Failure -> retry
            if attempt < max_retries:
                print(f"[Execution Failure] Timeout. Retrying {attempt+1}/{max_retries}...")
                continue
            return try_fallback_tool(args)  # fallback 툴로 전환
        
        except KeyError as e:
            # Specification Drift -> 런타임 스키마로 재매핑
            print(f"[Specification Drift] Missing field: {e}. Remapping...")
            args = remap_args_to_runtime_schema(args, e)
        
        except TypeError as e:
            # Invocation Error -> 인자 구조 수정
            print(f"[Invocation Error] Argument error: {e}. Repairing args...")
            args = repair_invocation_args(args, e)
    
    return None


def cross_source_verify(results: list):
    """
    Cross-source Conflict: 여러 툴 결과를 교차 검증.
    단순 다수결 X -> 원본 기준값과 대조 후 정규화된 값만 채택.
    """
    normalized = [normalize_value(r) for r in results if r is not None]
    
    # 일치하는 값만 신뢰
    from collections import Counter
    counts = Counter(normalized)
    most_common_val, count = counts.most_common(1)[0]
    
    if count >= len(results) // 2 + 1:  # 과반수 동의
        return most_common_val
    else:
        # 충돌 해결 불가 -> 추가 검증 요청
        raise ValueError(f"Cross-source conflict unresolved: {counts}")


# 에이전트 정책 프롬프트 (논문 Table 13 기반)
POLICY_PROMPT = """
You are a tool-orchestration policy model.

Rules:
- If a tool fails, diagnose WHY before retrying:
  * TimeoutError -> retry once, then fallback
  * Missing field -> remap to observed runtime schema  
  * Argument error -> repair argument structure
  * Conflicting results -> cross-check all sources before finalizing
- Finish ONLY when all required evidence branches are resolved.
- Treat 'success=false', missing fields, and errors as UNRESOLVED.
- Do not mistake partial results for final answers.

Return JSON: {"action": "call_tool|retry|fallback|finish", "reason": "..."}
"""