Financial agent architecture separates research from permissions and execution
A Juejin AI article proposes a financial-agent architecture that keeps model-led research distinct from data retrieval, deterministic computation, access controls, human review, and business execution. It uses a synthetic fund-portfolio case to show how portfolio questions can progress from public research to account access and proposed operations, while preserving share class, reporting period, comparability, calculation context, and risk concerns. The article says ETF Agent's current conversational research layer does not trade directly; orders, settlement, and institutional approvals are target divisions for future business-system integration.
The article offers an implementation-oriented framework for connecting agentic financial research with controlled operations without treating a model response as both investment advice and transaction authorization.
用户问：“我已经买了两只科技基金，再加一只，是增加分散，还是把同一种风险又买了一遍？”
模型可以解释分散投资，也可以调用工具查询基金、阅读持仓、计算重合。但这些动作做过一遍，并不意味着问题已经回答。系统还需要知道：读的是哪类基金份额，持仓属于哪一期，两份数据能不能比较，计算结果意味着什么，以及用户究竟在担心哪一种风险。
用户接着说：“那就按你说的调整。”事情又变了。前面是一项研究，现在可能涉及账户与资金。谁核对产品和金额？谁批准操作？没有收到明确结果时，应该查询，还是再提交一次？
金融 Agent 的全景，不是把更多模型、工具和审核模块放在一起，而是把研究与业务接好：模型可以灵活探索，数据、计算、权限和真实操作分别有人负责。 只追求自主性，会忽略行为后果；只追求控制，又可能让用户始终看不到有价值的结果。