Team D: 라우팅 로직

분석 대상

Collaborator Agent의 시스템 프롬프트 (라우팅 규칙), send_message_tool 구현

라우팅 시스템 프롬프트 핵심

def get_default_system_prompt(self):
    return f"""
    Role: Lead NetOps Orchestrator

    Specialist Agents:
    1. Connectivity_Troubleshooting_Agent — DNS, 연결성, 보안 그룹
    2. Performance_Agent — Flow Monitor, 트래픽 미러링, TCP 분석

    Route to Connectivity_Troubleshooting_Agent for:
    - Connectivity checks between hosts
    - DNS resolution issues
    - Security group and firewall diagnostics
    - Cannot connect or cannot reach errors

    Route to Performance_Agent for:
    - Network performance analysis
    - VPC Flow Logs monitoring
    - Traffic mirroring and PCAP analysis
    - TCP retransmission issues
    - Slow connection or high latency

    Agent Naming: Always use exact names:
    - "Connectivity_Troubleshooting_Agent"
    - "Performance_Agent"

    <Available Agents> {available_agents} </Available Agents>
    """

라우팅 결정 과정

사용자: "Fix retransmission issues for ExampleCorp"
    ↓
LLM 추론:
    1. "retransmission" → Performance 키워드 매칭
    2. 라우팅 결정: Performance_Agent
    3. send_message_tool 호출 결정
    ↓
도구 호출:
    send_message_tool(
        agent_name="Performance_Agent",
        task="Fix retransmission issues for ExampleCorp Image Application"
    )

send_message_tool 구현 상세

@tool
async def send_message_tool(agent_name: str, task: str) -> str:
    """Tool function for sending messages to remote agents."""
    return await self._send_message_impl(agent_name, task)

async def _send_message_impl(self, agent_name, task):
    # 1. 중복 방지
    request_key = f"{agent_name}:{hash(task)}"
    if request_key in self.completed_requests:
        return "Request already processed"

    # 2. A2A 메시지 구성
    payload = {
        "message": {
            "role": "user",
            "parts": [{"type": "text", "text": task}],
            "messageId": str(uuid.uuid4()),
            "contextId": str(uuid.uuid4()),
        }
    }

    # 3. SendMessageRequest 전송
    message_request = SendMessageRequest(
        id=message_id,
        params=MessageSendParams.model_validate(payload)
    )
    send_response = await client.send_message(message_request)

    # 4. 응답 파싱 (artifacts → parts → text)
    json_content = json.loads(send_response.root.model_dump_json())
    resp = []
    for artifact in json_content["result"]["artifacts"]:
        resp.extend(artifact["parts"])
    return json.dumps(resp, indent=2)

키워드 → 에이전트 매핑

사용자 키워드	라우팅 대상	근거
"check connectivity"	Connectivity Agent	"Connectivity checks between hosts"
"DNS resolution"	Connectivity Agent	"DNS resolution issues"
"cannot connect"	Connectivity Agent	"Cannot connect errors"
"retransmission"	Performance Agent	"TCP retransmission issues"
"slow connection"	Performance Agent	"Slow connection or high latency"
"traffic mirroring"	Performance Agent	"Traffic mirroring and PCAP"
"flow monitor"	Performance Agent	"VPC Flow Logs monitoring"

결과 전달 규칙

Communication Style:
- Be direct and technical
- Use bullet points for clarity
- CRITICAL: Do NOT use emojis
- Always relay complete analysis results from specialist agents
- Do not ask for permission before contacting specialist agents

CRITICAL for Performance Agent results:
1. PCAP 파일명 정확히 전달
2. tshark 분석 결과 상세 포함
3. S3 경로 포함
4. 요약하거나 생략 금지

에러 처리

# A2A 통신 실패 시
max_retries = 1  # 최소 재시도 (중복 방지)

for retry_count in range(max_retries + 1):
    try:
        send_response = await client.send_message(message_request)
        return parse_response(send_response)
    except Exception as e:
        if "throttling" in str(e).lower() and retry_count < max_retries:
            await asyncio.sleep(2.0 * (2 ** retry_count))
            continue
        return f"Error sending message to {agent_name}: {str(e)}"

토론 질문

1. LLM이 잘못된 에이전트로 라우팅하면 어떻게 되나요? 복구 메커니즘이 있나요?

2. 시스템 프롬프트의 키워드 매핑만으로 충분한가요? Agent Card의 tags를 활용할 수는 없나요?

3. send_message_tool이 하나의 도구인 이유는? 에이전트별 도구를 만들 수도 있지 않나요?

4. completed_requests의 키가 hash(task)를 사용하는데, 해시 충돌 가능성은?