使用说明
- 结构：一个主代理、协调器、四个工作代理
- 保持所有代理的状态结构相同，并将此状态存储在编排器中，以便所有代理都可以对其执行读写操作。此状态必须覆盖所有字段。
- 首先，主代理启动对话，并在状态中填写customer_id字段，同时执行常规聊天机器人操作。
- 销售代理：此customer_id随后由销售代理获取，并使用mcp服务器中的get_customer_info工具获取并写入编排状态中所需的协商详细信息。
- 然后，主代理寻求用户的批准，并在状态字段中更新该批准。如果状态不确定，协调器将返回销售代理以更新谈判条款。
- 一旦用户确认编排状态中的状态字段，编排器就会移动到验证代理。
- 验证代理：验证代理使用mcp服务器的verify_kyc工具验证用户，并在编排共享状态下更新用户。
- 一旦验证，流程将进一步转移到承销代理。
- 承销代理：承销代理使用get_credit_score和upload_salary_slip等工具获取详细信息并执行工具功能中提到的操作。
- 根据报价的结果和可接受性，主代理在收到制裁函生成器代理开始工作的确认后，寻求客户/用户最终接受贷款。
- 制裁函生成器：在获得完全批准后，它使用generate_Sanction_Letter mcp工具生成pdf。
- 最终，主代理表示祝贺，并向用户发送了批准函
共享状态：
从键入导入TypedDict、Dict、Any、Optional、List
类编排状态：
# -------------------------------------------------------
# 1. MASTER AGENT / USER INPUT
# -------------------------------------------------------
customer_id: str # Filled by master agent when user provides ID
user_query: str # Raw user input for conversation continuity
status: str # negotiation_status: "pending" | "approved" | "rejected"
# -------------------------------------------------------
# 2. CUSTOMER DATA (from MCP: get_customer_info)
# -------------------------------------------------------
customer_info: Dict[str, Any] # Full customer dict returned by MCP server
# -------------------------------------------------------
# 3. NEGOTIATION (Sales Agent)
# -------------------------------------------------------
requested_amount: Optional[int] # Customer-requested amount
preferred_tenure_months: Optional[int] # Customer preferred tenure
max_interest_rate: Optional[float] # User threshold
negotiated_offer: Dict[str, Any] # {
# "customer_id": str,
# "approved_amount": int,
# "tenure_months": int,
# "interest_rate": float,
# "justification": str
# }
# -------------------------------------------------------
# 4. KYC VERIFICATION (Verification Agent)
# -------------------------------------------------------
kyc_result: Dict[str, Any] # {
# "phone_verified": bool,
# "address_verified": bool
# }
kyc_status: str # "pending" | "verified" | "failed"
# -------------------------------------------------------
# 5. UNDERWRITING (Underwriting Agent)
# -------------------------------------------------------
credit_score: Optional[int] # fetched from get_credit_score
salary_slip_resource: Optional[str] # resource://<file> link after upload
underwriting_input: Dict[str, Any] # the exact MCP input payload
underwriting_result: Dict[str, Any] # MCP underwriting output
underwriting_status: str # "pending" | "approved" | "rejected" | "need_documents"
# -------------------------------------------------------
# 6. SANCTION LETTER (Sanction Letter Generator Agent)
# -------------------------------------------------------
sanction_letter_resource: Optional[str] # resource://<pdf>
sanction_letter_path: Optional[str] # local server path for debugging
sanction_letter_status: str # "generated" or "failed"
# -------------------------------------------------------
# 7. SYSTEM FLOW CONTROL (used by orchestrator)
# -------------------------------------------------------
flow_stage: str # one of:
# "start"
# "collect_customer_id"
# "negotiation"
# "await_user_confirmation"
# "verification"
# "underwriting"
# "sanction_generation"
# "complete"
messages: List[Dict[str, Any]] # optional conversation history for master agent
# -------------------------------------------------------
# 8. LOGGING / AUDIT
# -------------------------------------------------------
log_events: List[Dict[str, Any]] # track events locally before sending to MCP log_event
