` 同步校注（2026-11-04）：文中涉及撤权、来源消费和恢复的历史案例，应与当前 Run 授权生命周期区分。当前实现中，Host 已授予的权限在同一 Run 及其恢复中持续有效，不因外部权限表变化重新撤销；新 Run 独立准入。账户、scope、来源范围、内容身份、预算和显式取消仍按原合同执行。历史撤权复现用于解释当时的问题，不代表当前 Run 内重新授权机制。
原则篇｜Pi、Claude Agent SDK、LangGraph 与金融 Agent 的职责、取舍和退出条件。
当一个 Agent 已经能够搜索网页、阅读报告、执行 Python、调用子代理、保存会话，我们很容易产生一个判断：再做金融垂类框架，是否只是在成熟工具外面换了一层名字？
这个怀疑是合理的。一个系统接上行情接口、加几段金融 Prompt、注册几个基金 Skill，并不因此拥有新的架构价值。反过来，金融业务确实存在专门的事实、计算和责任问题，也不能因为通用框架功能丰富，就把这些问题当成模型会自动处理的细节。
本文的判断是：金融领域能力通常需要建设，但独立的金融 Agent 框架并不总是需要，自研执行运行时更不是顺理成章的下一步。 必须分别证明这三个命题，不能从“基金口径很复杂”直接跳到“所以应该重写 Agent Loop”。
下面沿一项基金组合研究展开。我们会看到金融规则怎样改变计算，证据怎样限制结论，模型自主性应当放在哪里，以及怎样选择最小可用方案。文中的资金、产品和日期组合均为教学构造，不是作者或任何客户的实际持仓；关于框架的能力描述依据文末官方文档，核对日期为 2026 年 11 月 2 日。
一、先把问题说完整：我们究竟在比较什么？
- 四个经常被混为一谈的东西 模型、执行框架、领域能力和业务产品不是同一层对象。模型负责解释问题与生成行动；执行框架承接模型、工具、上下文和任务状态；领域能力定义基金身份、收益口径等业务含义；产品还需要账户、权限、报告、界面、运维及用户实际使用过程。 这不是要求工程代码必须分成四个目录，更不是建议增加四个服务。它只是帮助判断：某个困难究竟应该在哪一处解决。一个应用完全可以在同一进程中实现这些职责；相同职责已经被现有模块清晰表达时，不必再造一层抽象。 如果把完整金融产品与一个通用 SDK 放在同一张功能数量表里，产品通常会显得“什么都有”。但这种比较把产品自己接入的数据、业务流程和界面，全部算成了执行框架的优势，实际上没有回答应该选哪套基础设施。
- Pi、Claude Agent SDK、LangGraph 各自解决什么
Pi 官方把自己定位为可扩展的轻量 Agent Harness，也就是模型工作的执行环境。它提供会话、工具与上下文等机制，并通过扩展、Skill、SDK 和 RPC 让开发者接入自己的流程。其官方说明刻意不把 MCP、子代理和权限弹窗全部做成内建功能，而是允许按环境补充；这是设计取舍，不应自动被评价为能力缺陷。[1]
Claude Agent SDK 提供 Claude Code 的 Agent Loop、工具与上下文管理，支持 Python 和 TypeScript。它还提供 MCP、权限、Hook、子代理、会话继续等能力。因此，“我们需要工具、恢复和专家分工”本身并不能构成重新开发运行时的充分理由。[2]
LangGraph 是面向有状态、长时任务的底层编排框架与运行时。确定性代码和模型驱动步骤可以在同一个图中组合，持久化与人工中断是其正式能力。把它简单描述成“只能按固定流程执行”，同样不准确；图可以编排自主研究，而不只是线性的业务审批。[3]
三者不是彼此完全可替代的商品，更没有脱离任务的唯一冠军。比较时要问：我们需要接管多少执行语义，愿意承担多少集成工作，哪些状态必须跨进程保留，哪些工具和部署环境已经由现有系统提供。
图 1：领域系统与通用执行能力的关系
领域目标驱动执行 — 读取和计算也会反过来更新业务条件
图中回到领域目标的箭头很重要：研究不是把预先排好的工具执行一遍。新资料可能证明最初问题需要改写，或者只能回答其中一部分。执行框架使这种循环能够发生，领域能力帮助解释每一次观察意味着什么。
二、用同一个组合，区分“回答了”与“研究完成了”
- 教学案例：三只基金、一笔新增资金 假设一个模拟组合总额为 120,000 元。A 基金占 45%，B 基金占 30%，C 基金占 15%，现金占 10%。A、B 的产品资料都明确跟踪同一个指数；C 是主动管理的主题基金。用户另外计划投入 20,000 元，并提出一个完整问题： “我的基金是不是买重复了？新增资金怎样安排，才能降低集中度，又保留 10% 现金？请把不能确定的部分讲清楚，不要替我交易。” 这里同时存在事实识别、组合计算、条件比较和权限约束。简单回答“A、B 重复，卖掉一只”遗漏了新增资金、现金目标和不执行交易三个重要条件。相反，把所有条件都写成一大段免责声明，也没有完成研究。 合理的目标应当是：先确认已有配置的可证明特征，再判断新增资金能解决什么问题，明确需要哪些额外资料，最后交付有条件的方案比较。用户没有要求系统替他完成真实买卖，研究结果也不自动取得执行资格。
- 第一个可以立即成立的结论 只要账户快照和两只基金的跟踪基准得到确认，就可以说：总资产的 75% 配置在跟踪相同指数的两只基金上。这个结论来自 45% 加 30%，属于基准层面的重复配置，不需要先穿透全部底层证券。 但它不能自动升级为“75% 资产完全重合”，更不能推出应该卖出其中某只。现金比例、复制方式、实际组合、费用和用户原有持有条件都可能影响后续判断。正确的方法是保留已成立的结论，而不是因为更深层资料缺失就把整题拒绝。 这也是金融 Agent 很容易走偏的地方：为了追求“完整”，它可能把未知部分补成一个精确数字；为了追求“安全”，又可能连已经足够支持的事实都不愿输出。两种极端都损害实用性。
- 为什么这里需要 Agent，而不只是计算器
如果任务只有“45% 加 30% 等于多少”，确定性代码足够。如果任务要求根据资料可得性选择研究层次，发现 C 的披露不足，寻找可替代来源，比较降低集中度的不同路径，并据此调整问题，模型驱动的研究循环才有实际价值。
因此，领域 Agent 的独特作用不在于重新发明加法，而在于围绕目标组织证据与方法。确定性程序承接可以精确定义的运算，模型承担开放式研究和解释；两者的结合应当产生比固定模板更有用的结果，而不是更多工具调用记录。
三、金融事实的最小单位，不是一个裸数字
- 数字离开对象、单位与分母，通常没有足够意义 假设一个工具返回“权重 8%”。我们仍然不知道它是谁对谁的权重：某只基金中的某家公司，某个账户中的某只基金，还是研究样本中的行业分类？分母可能是基金净资产、总资产、股票资产，也可能只是已经披露的前十大持仓。 即使所有数值都来自真实来源，只要分母错配，最终计算仍可能错误。最危险的情况不是数值缺失，而是数据看起来齐全，模型也能顺利算出结果，但不同字段实际上回答的是不同问题。 我们可以用“带语义的事实”代替裸数字：对象是什么，指标是什么，原值是什么，单位和尺度是什么，分母是什么，适用时点是什么，来源在哪里。不是每个字段都要复制进每个 JSON，但调用方必须能够通过明确的关系取得这些信息。
- 产品身份应当服务于问题，而不是机械去重 同一基金的不同份额类别，可能共享部分投资组合信息，同时具有不同费用或分配安排。分析底层资产时可能需要识别共同组合，计算持有成本时又必须保留具体份额、渠道和适用条款。不能选一个“统一基金 ID”，就认为所有层次都可以合并。 类似地，两只基金跟踪同一指数，不等于它们是同一投资产品；两条账户记录对应同一只基金，也不等于其中一条应该被删除。去重的是研究资料、产品实体还是资金持有记录，必须先说清楚，否则去重可能把真实资本一起去掉。 这类区分属于领域语义，而不专属于 Agent。已有金融数据服务已经维护得当时，直接复用比重新创建 Agent 专属标的中心更合理。领域框架应减少调用方必须理解的细节，而不是再复制一份权威数据库。
- 类型化只能阻断部分错误
把金额定义成带币种的十进制字符串，可以阻止 CNY 与 USD 被无意相加；把份额和净值区分为不同量，可以帮助检查乘法的单位关系。但正确的类型不能证明基金身份正确，也不能证明净值来自适用日期。
因此，类型系统、运行时校验、来源核对和业务解释是互补手段。任何一个都不应该被包装成“金融正确性已经解决”。尤其不能因为一个结果通过了结构化 Schema，就给整篇报告盖上可信认证。
四、时间不是一个过期时间，而是一组不同的问题
- 数据描述的时间，与我们知道它的时间 假设一份教学报告描述 9 月 30 日的持仓，10 月 25 日披露，Agent 在 11 月 2 日读取。三个日期都是真的，但作用不同：第一个描述资产状态，第二个描述信息进入公开范围的时间，第三个描述本次系统取得资料的时间。 “今天取得的最新报告”不能被简写成“今天的持仓”。反过来，历史持仓也不因为不是今天的就失去价值；它可以用于解释当期配置、核对过去判断或构建有明确时点的研究。有效性取决于任务，而不是一刀切的天数门槛。 如果要回测 10 月 1 日能否作出某项判断，那份 10 月 25 日才披露的报告不能被当成当时已知信息。这个例子说明，保存数据截止日还不够，有时必须同时保存披露或可获知时间，才能避免把后来信息带回过去。 图 2：同一份资料的三个时间
三个时间，两种研究问题 — 教学日期；同一资料不等于同一适用范围 打开完整矢量图
- 新资料出现时，不要悄悄改写旧结论 假设用户阅读报告后又导入一份纠正版持仓。新版本应该影响后续计算，但已经发布的旧报告也有自己的来源和历史意义。把底层记录直接覆盖，然后让旧报告自动显示新数字，会使用户无法知道自己当时看到的究竟是什么。 更合理的设计是把事实修订和结果修订连接起来：旧结果仍绑定旧输入，新结果绑定新输入；展示时解释两者的差异。需要重新计算的范围由真实依赖决定，而不是发现一个来源变化就把所有历史研究全部重跑。 这里的“不可变”也不是永久保留一切的命令。保存策略仍需遵守权限、数据最小化和适用留存规则。版本身份解决的是如何辨认不同事实，不等于为无限期保存私人资料提供理由。
- 时间策略也要允许退出
一次性阅读公开说明书时，记录来源和版本可能已经足够，不必构建双时间数据库。只有当任务需要历史可得性、跨期观点比较或修订追踪时，额外时间语义才值得进入执行合同。
这条原则具有普遍性：解释时能够区分的概念，并不都要立即成为系统字段。先确认区分会改变哪个判断、哪项计算或哪条恢复行为，再决定是否增加结构。否则领域建模会从解决真实问题变成维护百科全书。
五、一个分母错误，足以让“精确穿透”失去意义
- 把教学组合穿透到公司 X 继续使用前述组合。假设在已经确认的同口径资料中，A 对公司 X 的权重为 8%，B 为 10%，C 为 6%。公司 X 占总资产的已知暴露为 45%×8%＋30%×10%＋15%×6%，即 7.5%，对应 9,000 元。 这只是一项教学计算。它成立的前提包括基金权重和底层权重的定义一致、输入适用于所声明的时间、公司身份映射正确，以及不存在未经处理的重复层级。公式简单，不代表这些前提也简单。 再假设 A、B、C 可取得的底层披露覆盖分别为 60%、70%、40%。按总资产加权，已披露基金底层覆盖为 54%。现金另占 10%，剩余 36% 总资产的基金底层去向尚未由这些资料解释。
- 三个看起来都合理的百分比 公司 X 占总资产为 7.5%；占基金资产为 7.5%÷90%，约 8.33%；占已披露基金底层样本为 7.5%÷54%，约 13.89%。三个算式都可以算对，但回答不同问题。 假如系统把第三个数放在“组合实际公司权重”栏中，就发生了样本归一化误用。问题不在小数精度，而在把已知样本的结构当成整个组合的结构。增加更多小数位只会让错误显得更有权威。 资料缺失也不能被填成零。零意味着我们知道没有暴露，未知意味着我们还不知道是否存在暴露。只有明确满足非负、无杠杆等假设时，才能把已知部分解释为相应下界；不能把这种教学假设推广到含衍生品、空头或净敞口的全部基金。 图 3：同一数值，三个不同分母
同一金额，三个分母 — 教学例：公司 X 已知金额 9,000 元 打开完整矢量图
- 现金目标为什么必须使用调整后的分母
原组合现金为 12,000 元。新增 20,000 元后，总资产在不考虑市场变化、费用和其他现金流的教学假设下变成 140,000 元。保留 10% 现金意味着最终现金应为 14,000 元，因此新增资金中可投入基金的是 18,000 元，而不是全部 20,000 元。
如果系统仍按原来的 120,000 元计算目标现金，就会把 12,000 元当成足额储备。这种错误不需要复杂金融知识，只需要分清“当前组合”和“调整后组合”。但在多轮对话中，新增金额、目标比例和旧快照很容易被混用。
这也是领域方法值得封装的原因：同一现金守恒与分母规则可能被问诊、再平衡、定投比较和周报重复使用。让每次模型临时生成计算步骤，难以保证这些入口长期保持相同含义。
六、确定性计算也可能使用了错误的方法
- 正确执行公式，不等于公式适用于问题 金融 Agent 经常把 Python 当成可靠性的分水岭：只要调用代码，数字就可信了。实际上，代码可以准确执行一个错误的公式，也可以对错误的数据对象做精确计算。我们需要区分算术正确、方法适用、输入真实和结论合理四件事。 例如，某个教学基金单位净值从 1.00 变成 0.99，期间每份现金分配 0.04，且假设分配不再投资、没有其他现金流和费用调整。单期持有总回报为 3%，不是净值变化得到的负 1%。如果工具只计算净值差，即使使用无限精度也仍然答错了问题。 若分配再投资，还要明确再投资日期、价格和份额变化；存在外部资金流时，又要选择与问题匹配的收益度量。这里不需要为了文章列尽所有金融公式，只需要理解：算法名称、输入口径和适用条件是结果的一部分。
- 费用需要区分假设、条款估算和实际发生 “假设费率为某个值”“依据当前适用产品条款估算费用”“账户回执显示实际扣费”是三种不同事实。第一种可以用于敏感性分析，第二种需要具体份额、渠道、持有期及有效日期，第三种需要真实执行证据。 它们不应共用一个不加说明的“手续费”字段。尤其是某些费用已经反映在净值中时，再从净值收益扣除一次可能造成重复计算；是否需要调整，应由明确的方法合同决定，而不是由模型看到“费率”两个字就自动减去。 金融规则会变化，产品条款也会修订。因此，所谓领域不变量不是“某个费率永远固定”，而是“任何费用估算都必须知道适用的条款版本与计费基础”。稳定的是判断所需的关系，不一定是关系中的参数。
- 不要为统一而建立另一个金融计算中心
当不同模块已经拥有明确的方法时，统一的重点应是共同语义与结果接口，而不是把所有公式强迁移到一个全局 Manager。组合现金、收益归因、费用估算和底层暴露可以分别由最合适的维护方负责。
值得共享的可能是单位运算、十进制规则、输入引用和版本约定；不同收益窗口、费用顺序和统计方法则应该保留差异。真正重复的公式需要合并，含义不同但名称相近的方法不应该为了目录整齐而合并。
这种联邦式设计允许领域能力逐步成熟。只有一个消费者时，先在原模块修正错误；多个任务反复遇到同一问题时，再抽出深一些的公共接口。抽象应该减少调用方知识，而不是把原复杂度换成更长的配置表。
七、证据链应当约束主张，但不能冒充真值机器
- 至少区分来源、计算与解释 基金报告中的一行数据是来源事实，程序根据数据计算出的集中度是派生结果，模型据此解释“组合对某种风险较敏感”是研究判断。三者都重要，但它们具有不同的证据身份。 子代理写出的总结也不自动成为新的原始来源。总结可以帮助主代理定位问题，但主代理需要表达关键金额或引用具体表格时，应能回到对应版本的原材料或确定性结果。不能仅仅给专家回答分配一个 ID，就让它拥有与正式披露相同的事实地位。 建立这种关系并不要求每句话都附一个巨大的证据对象。关键是保留能够支撑重要结论的最小链路：哪份资料、哪个位置、什么输入、什么方法、哪个结果，以及结论实际使用了哪一部分。 图 4：证据支持主张，不自动认证主张
证据支持，不自动认证 — 来源、计算和研究主张保留不同身份 打开完整矢量图
- 有链接、哈希和血缘，仍可能答错 一个链接可以指向真实网页，却不支持旁边那句话。一个哈希可以证明保存的字节没有变化，却不能证明这些字节来自正确产品。数据血缘可以完整记录计算路径，却仍然完整地记录了一次错误方法。 所以，证据系统首先提供的是可检查性，而不是正确性保证。它让人能够追问、复算和定位问题，减少错误隐藏在自然语言中的机会。把这种工程价值说清楚，比声称“全链路可信，因此研究已经可靠”更准确。 独立复核也需要独立条件。第二个模型如果只阅读第一个模型写好的摘要，可能只是重复相同错误。更有意义的复核应在必要时取得原始依据、独立算式和不同方向的反证，而不是给流畅程度打分。
- 保留局部结论，比追求全局可信布尔值更有用
在组合案例中，A、B 的基准重复可以确认，C 的完整底层暴露不能确认，新增资金后的现金算式又可以独立成立。这些结论应当各守范围，而不是被压成一个“报告可信／不可信”的总开关。
实际输出可以清楚区分“已确认”“基于显式假设的情景”“仍缺资料”。必要输入不足时，相关计算不应生成虚假结果；但可选的额外检查失败，也不应自动撤销独立成立的事实。局部失败和整体交付应按真实依赖处理。
这不是放松金融质量，而是避免把质量治理做成一个过度阻断系统。用户真正需要的是知道哪些判断可以使用、哪些还需要补证，而不是看到一个看似严格、实际上解释不了原因的红色标签。
八、自主性要放在研究策略里，不能放在授权事实里
- 模型可以决定下一步研究什么 发现 C 的持仓资料不完整后，模型可以查找基金管理人报告、比较不同披露窗口，或者判断目前只能做策略层面的风险解释。它也可以提出一个更适合的问题：先识别同基准重复，再讨论新增资金是否能降低已知集中度。 这种策略调整正是 Agent 的价值。若系统把“先查 A、再查 B、再查 C、必须得出精确重合率”写死，即使模型已经发现问题无法按原口径回答，也只能继续完成一个错误计划。 但研究策略的灵活性不意味着所有规则都由模型决定。当前用户身份、可读账户、允许发送资料的供应商、投资账本和已经授予的权限，不应因模型认为“这样更方便”而改变。
- 研究、建议和执行需要区分，但不是每次都加审批 一个公开基金说明书摘要不需要先弹出交易授权框。一个只读组合分析也不应该因为出现“降低集中度”就自动启动全套交易流程。控制应该与实际动作和后果匹配，而不是与某个金融关键词绑定。 当任务涉及客户私有资料，需要检查数据访问和使用范围；当任务请求真实账户写入，需要核对具体动作的授权与有效条件。这里讨论的是工程责任分离，不是替任何地区、机构和服务类型出具法律适用结论。 “用户说了可以”还必须对应具体对象和版本。批准查看某份持仓，不等于批准发送给所有外部模型；批准某个金额的操作，也不等于后续金额改变仍继续有效。授权与动作的绑定比审批弹窗本身更重要。 图 5：探索可以开放，实际效果需要明确授权
开放探索，明确执行授权 — 研究自主性不等于外部写入权限 打开完整矢量图
开放探索，明确执行授权 — 研究自主性不等于外部写入权限 打开完整矢量图
- Prompt 是方法说明，不是安全边界
告诉模型“只读”“不要交易”“不要泄露账户”很有帮助，但如果底层工具仍提供任意数据库写入或无限制文件读取，真正的权限边界并没有建立。模型偏离指令时，系统应当仍能限制它造成的效果。
另一方面，也不应该把所有判断都改成机械规则。比如“这份研究是否有说服力”无法被一个固定关键词检查替代。确定性程序处理可定义的授权、身份和算术；语义质量通过适当评测与复核处理，二者不能相互冒充。
九、持久化保存的是执行状态，研究连续性还需要额外含义
- 恢复一个会话，不等于恢复一个研究问题 通用框架已经提供丰富的会话或持久化机制。LangGraph 明确区分线程内的 Checkpoint 与跨线程存储；Claude Agent SDK 也支持会话继续与分支。这些能力值得复用，而不是被垂类系统当成不存在。[2][4] 金融研究的额外问题是：恢复时仍然在回答同一个问题吗？使用哪一期资料？用户是否改变了目标？已经发布的阶段结论应保留还是修订？原来的数据授权是否仍然有效？这些含义需要应用或领域逻辑明确表达。 一个会话记录可以保存全部对话，却仍然无法区分“上周的假设”和“本周的事实”。因此，长期研究不只是更长的聊天历史，而是能够关联目标、依据、结论及其变化原因。
- 不要把恢复理解成盲目重复动作 研究读取失败后重新获取资料，和对外提交指令后没有收到回执，是不同问题。后者可能已经产生效果，简单重试可能造成重复。即使是保存报告，也要区分写入前失败、写入成功但确认丢失、状态暂时未知。 LangGraph 的中断文档明确提醒，恢复时中断之前的节点代码可能重新执行，因此相关副作用需要幂等设计。这个问题不是某个框架的缺陷，而是应用需要与运行时共同承担的执行语义。[5] 金融垂类系统的贡献，应当是给具体业务动作规定身份、重试和核对方式，而不是宣称拥有一个通用的“恰好一次”开关。外部系统没有相应协议时，本地数据库保存一个成功标记并不能创造外部恰好一次保证。
- 很多任务不需要长期记忆
一次性比较两份公开资料时，不保存长期偏好、不构建跨周研究工作区，可能是最好的设计。长期记忆会带来过期、冲突、权限及删除语义，应该由真实持续任务驱动。
需要持续追踪时，也优先保存有价值的稳定偏好和研究目标，而不是把每次生成的意见都沉淀成事实。过去模型说过“这只基金适合你”，不能在后续会话里无条件升级成用户已经确认的长期投资偏好。
十、领域能力、Skill 与框架之间，存在一道经济边界
- 可以先用 Skill 解决的问题，不必立刻工程化成平台 如果某个任务主要需要一套研究步骤，例如先确认基金身份，再找报告，再选择适用指标，那么 Skill 是合适的表达。它能够让模型按任务加载方法，不必把全部领域知识塞进常驻系统提示。 如果任务的关键是一个重复使用的确定性公式，就把公式放在工具或领域模块中。工具需要的参数、单位和失败条件应当清楚，Skill 负责说明什么时候使用它。不要用一篇更长的 Skill 去代替已经能够程序化表达的计算规则。 当多个任务共享标的身份、时点口径、来源引用和结果结构，并且这些规则需要独立演进与验证时，才开始形成领域 SDK 或框架。框架是复用需求的结果，不应该成为项目启动时必须填满的目录模板。
- 框架的价值，可以用删除实验来检验 设想删除某个公共模块。如果相关复杂度也一起消失，调用方更容易理解，它可能只是无价值的转发层。如果删除后，相同的费用口径和来源校验重新散落到多个入口中，那么这个模块确实在集中管理复杂度。 这个实验比“代码是否有一层 Interface”更有意义。接口的深度来自它隐藏了多少调用方本不该承担的细节，而不是类型名字有多抽象。一个只有一个适配器的抽象可能有用，但仍应承认跨运行时可移植性尚未被实际验证。 同样的判断也适用于独立部署。逻辑职责可以分离，却不必立即拆成微服务。只有部署隔离、资源调度或组织协作产生真实需求时，跨进程成本才可能值得承担。
- 正常任务也会为复杂架构付费
额外机制不触发时，是否仍增加上下文、装配步骤、数据库访问和故障面？新加一个工具，是否需要同步修改多份清单、恢复协议和测试？这些是框架的正常路径成本与变化放大成本。
本文不为它们编造一个统一百分比。不同部署和任务差异很大，应该用真实请求测量。但不能因为成本尚未量化就假定不存在，更不能只展示复杂机制在极端故障中的好处，而忽略普通查询每天都承担的额外负担。
因此，合理的扩展通常应当按需启用：短任务不启动长期工作区，没有独立子问题不启动专家，现有结果足够时不重复评审，可选增强不可用时不拖住独立合法的回答。
十一、最强的反对意见：直接把规则写进现有框架，不就够了吗？
- 是的，而且常常就是正确答案 用 Claude Agent SDK 运行研究、以 MCP 暴露基金工具，再加一个领域计算库，可以成为完整可用的方案。用 LangGraph 组织读取、计算和人工检查，或者用 Pi 扩展接入相同工具，也同样成立。 只要规则集中、测试明确、结果可核验，并不要求代码仓库里一定出现一个叫作 FinancialAgentFramework 的包。独立框架和合理业务模块之间不存在天然的价值等级，决定因素是实际复用与维护效果。 所以，本文并不试图证明“所有金融 Agent 都要自研框架”。恰恰相反：领域复杂度证明需要领域建模，不自动证明需要新的运行时。 谁主张增加一套执行系统，谁就需要补充执行层的独立理由。
- 哪些情况才可能支持自建运行时 当已有运行时无法在实际部署条件下满足某项关键执行合同，扩展点不足，兼容层反而比局部自研更复杂，才值得认真比较。可能的条件包括特殊的模型协议、资源准入、跨进程恢复或不可替代的交付语义，但每一项都需要具体反例支持。 “我们希望控制得更细”不够具体。应当说清哪个动作、哪种故障、现有框架怎样处理、为什么不符合业务要求，以及是否已经尝试较小的扩展。否则所谓控制力只是把成熟框架的维护责任搬到了自己团队。 即使最终自建，也应把替换成本算进去：协议流式分片、取消竞态、重试、插件生命周期和状态迁移都可能长期消耗工程时间。这些工作可能有研究价值，但不能自动记为金融研究差异化。 图 6：从最小方案出发，而不是从平台出发
先复用，再证明增量必要 — 领域核心与执行运行时是两种判断 打开完整矢量图
先复用，再证明增量必要 — 领域核心与执行运行时是两种判断 打开完整矢量图
- 一个反向的成功标准
如果迁移到更成熟的通用框架后，能删除大量自有执行代码，同时保持基金计算、证据与权限行为不变，这应当被视为成功，而不是技术能力退步。真正积累下来的领域能力不应依赖维持自研代码的规模。
相反，如果每次更换模型都需要重写基金口径，说明领域语义与执行细节耦合过深。垂类核心的价值在于稳定可复用的判断关系，而不是必须永远绑定某一代 Harness。
十二、多 Agent 的必要性，也需要单独证明
- 多角色不等于多份独立信息 把同一份资料交给“多头专家”“空头专家”“风险专家”，可能产生不同措辞，也可能产生有用反证，但它并不保证信息真的独立。三个模型都依据同一错误摘要，最终投票只会增加错误的声势。 子代理最明确的工程价值通常是隔离上下文、分解可并行的子问题、使用特定工具与方法。Claude Agent SDK 官方也把这些列为子代理的重要用途。是否改善金融判断，则还需要在具体任务中测量。[6] 例如，一名专家负责核对披露时点，另一名负责复算现金与费用，第三名寻找推翻配置假设的证据，这比单纯给三个人格起名字更接近有用分工。但若主代理已经能低成本完成这些检查，分工也可能没有净收益。
- 主代理不能只收取“专家说没问题” 有价值的交接应当包含专家回答了什么子问题、依据什么资料、计算采用什么方法、仍有什么缺口。主代理需要知道哪些是原始事实、哪些是专家意见，必要时可以回读关键原材料。 否则，我们只是在主代理与事实之间增加了一个无法检查的自然语言代理层。上下文虽然变短，研究责任却变模糊。专家数量越多，资料权限、预算、取消和恢复的复杂度也越高。
- 把退出条件写进设计
没有独立子问题时不启动专家；资料不足时不靠更多角色弥补；评审没有发现额外错误且显著增加延迟时，回到单主代理基线。退出机制不是对高级架构缺乏信心，而是避免把某种组织形态变成必须长期支付的固定成本。
金融 Agent 的目标是更好地完成研究，不是展示一次任务能够同时运行多少 Agent。多代理改善了什么，需要与同题、同资料窗口、相近模型条件下的简单基线比较，不能只展示一次热闹的成功演示。
十三、评价体系要回答用户问题，而不只统计工具事件
- 执行、研究、交付和采用是不同观察面 一次运行结束，说明执行器进入终态；报告保存成功，说明某些字节已经落盘；用户能够下载，说明交付链路成立；结论经独立复核，才说明在相应范围内研究质量得到支持。这些事实相互关联，却不能互相替代。 同样，用户读懂并采用一份研究，也不直接证明它提高了投资收益。投资结果受到市场与行为等因素影响，不能从工程验收推出长期回报。把不同观察面分开，不是抹杀进展，而是让每一种进展都得到准确评价。 图 7：四种完成，不是一条自动升级链
四种完成，分别核对 — 观察面彼此相关，但不能自动升级 打开完整矢量图
- 评测应当包含能够改变结论的反例 组合案例至少可以加入这些条件变
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
`Sync Note (2026-11-04): Historical cases involving divestiture, source consumption and recovery should be distinguished from the current authorized life cycle of Run. At present, the permission granted by Host continues to be valid in the same Run and its recovery, without being revoked due to external permission table changes; the new Run has independent access. Accounts, scope, source range, content identity, budget and prominent cancellations are still performed under the original contract. Historical revocations are used to explain the problems at the time and do not represent the current reauthorization mechanism in Run.
The Principles cover the duties, trade-offs and exit conditions of Pi, Claude Agent SDK, LangGraph and Finance Agent.
When an Agent has been able to search the web page, read the report, execute Python, call a sub-agent, save a session, we can easily make a judgment: is it just a name changed outside the mature tool?
This suspicion is reasonable. A system connected to the interface, a few financial segments Prompt, the registration of several funds, Skill, does not have new structural value. In turn, financial operations do have specific factual, computational and liability issues, which should not be automatically addressed as models because of the extensive functions of the common framework.
The judgement here is that while capacity-building is usually required in the financial field, an independent financial Agent framework is not always needed, and self-study is not a logical next step when it comes to operating. It is important to prove separately that the three propositions cannot go directly from “the Fund's complex calibre” to “Agent Loop should therefore be rewritten”.
The following follows a portfolio study. We will see how financial rules change calculations, how evidence limits conclusions, where model autonomy should be placed, and how the least available options are chosen. The funds, products and date combinations in the text are the teaching architecture, not the actual hold of the author or any client; the competency description of the framework is based on the end-of-the-text official document and is checked on 2 November 2026.
First, complete the question: What are we comparing?
- Four things that are often confused with models, implementation frameworks, field capabilities and business products are not the same audience. The model is responsible for explaining problems and generating actions; implementing framework acceptance models, tools, context and mission status; defining business implications such as the identity of the fund, the calibre of proceeds, etc., in the field of competence; and requiring accounts, permission, reports, interfaces, traffic and the actual user process. This is not to require that the project code be divided into four directories, let alone to recommend four additional services. It merely helps to determine where a particular difficulty should be resolved. An application could well achieve these functions in the same process; when the same functions have been clearly expressed in the existing module, there is no need for an abstraction. If a complete financial product is placed in the same functional list as a common SDK, the product usually appears to be "everything." This comparison, however, fully captures the data, business processes and interfaces that the products themselves access as the advantages of the implementation framework and does not actually answer which infrastructure should be selected.
- Pi, Claude Agent SDK, LangGraph.
The official Pi has positioned itself as a light, scalable Agent Harness, the implementing environment for modelling. It provides mechanisms such as sessions, tools and context and allows developers access to their own processes through extensions, Skill, SDK and RPC. Its official statement deliberately did not allow for the internalization of all MCP, sub-agents and permission bullet windows, but rather allowed for environmental supplementation; it was a design trade-off and should not be automatically evaluated as a capability deficit. [1]
Claude Agent SDK supports Python and TypeScript by providing Claude Code's Agent Loop, tools and context management. It also provides MCP, permission, Hook, sub-agents, continuity of sessions, etc. Thus, “We need tools, rehabilitation and a division of labour among experts” cannot in itself constitute sufficient justification for re-development of operations. [2]
LangGraph is the bottom layout and running time for a current, long-term task. The certainty code and model-driven steps can be combined in the same picture, and sustainability and manual disruption are its formal capabilities. It is also not accurate to describe it simply as “implemented only by a fixed process”; it allows for the organization of autonomous studies rather than merely linear business approvals. [3]
The three are not completely substitutes for each other, let alone the only champions of the mission. The comparison asks how much implementation semantics we need to take over, how much integration we are willing to undertake, what status we have to keep across the process and which tools and deployment environments are already provided by existing systems.
Figure 1: System relationship to generic implementation capacity
Area-driven implementation - reading and computing will in turn update business conditions
It is important to return the arrows to the field objectives in the figure: research is not done through pre-packaged tools. The new information may prove that the initial questions need to be rewritten or that only part of them can be answered. The implementation framework enables such cycles to take place, and field capacity helps explain what each observation means.
II. Distinguishing between “responses” and “completed studies” using the same combination
- Teaching cases: three funds, one additional fund, assuming a simulation portfolio totalling $120,000. Fund A accounts for 45 per cent, Fund B 30 per cent, Fund C 15 per cent and Cash 10 per cent. The product information for A and B clearly tracks the same index; C is a voluntary thematic fund. The user plans to invest another $20,000 and asks the complete question: "Is my fund a duplicate? How can the additional funds be arranged to reduce concentration and retain 10% of cash? Please make clear what is not certain and do not trade for me.” There are both factual identification, grouping, comparison and permission constraints. The simple answer “A, B, repeat, sell one” omits three important conditions for new funds, cash targets and non-performance. Instead, all the conditions were written into a large disclaimer and the study was not completed. A reasonable objective should be to identify the probative features of the configuration, to judge what the additional funds would solve, to identify what additional information was needed, and to make final delivery conditional. The user did not require the system to complete the real deal for him, nor did the results of the study automatically qualify for execution.
- The first conclusion that can be established immediately is that, as long as a snapshot of the accounts and the tracking benchmarks of the two funds are confirmed, 75 per cent of total assets are allocated to two funds that track the same index. This conclusion comes from 45% plus 30%, which is a duplicate configuration at the baseline level and does not need to penetrate all bottom securities first. But it can't automatically be upgraded to "75% full overlap," let alone roll out should sell one of them. Cash ratios, reproduction patterns, actual portfolios, costs and user holdings may all affect subsequent judgement. The correct approach was to retain the established conclusion, rather than rejecting the whole issue because of a deeper lack of information. This is also where finance Agent can easily be biased: in the pursuit of “integrity”, it may add the unknown to an exact number; in the pursuit of “security”, it may not even be willing to export the facts that are already sufficiently supported. Both extremes undermine practicality.
- Why do you need Agent here, not just the calculator?
If the mission is only "45% plus 30% equals what," the certainty code is enough. The model-driven research cycle is of practical value only if the task requires the selection of the study level according to the availability of information, the discovery of insufficient disclosure of C, the search for alternative sources, and the adjustment of different paths to lower concentration.
Thus, the unique role of the field Agent is not to reinvent additions, but to organize evidence and methods around the target. The acceptance of a certainty procedure can be precisely defined as an operation, and the model bears open research and interpretation; a combination of the two should produce more useful results than a fixed template, rather than more tool call records.
The smallest unit of financial fact is not a nudity
- Numbers leave objects, units and denominators, usually not meaningful enough to assume that a tool returns " 8% weight." We still do not know who its weight is: a company in a fund, a fund in an account, or a study of industry classifications in sample. The denominator may be the Fund ' s net assets, total assets, equity assets, or just the top 10 positions already disclosed. Even if all values are derived from real sources, the final calculation may be incorrect as long as the denominator is wrong. The most dangerous situation is not the absence of values, but the fact that the data appear to be complete and that the model can successfully calculate the results, but different field actually answers different questions. We can replace the nudity figure with the term “facts with semantics”: what is the object, what is the indicator, what is the original value, what is the unit and scale, what is the denominator, what is the point of application, where is the source. Not every field is copied into every JSON, but the caller must be able to obtain this information through a clear relationship.
- Product identity should be at the service of the problem rather than the mechanical weighting of the different categories of shares of the same fund, which may share some portfolio information, with different costs or allocation arrangements. Common combinations may need to be identified in the analysis of the lower assets, and specific shares, channels and applicable clauses must be retained in the calculation of holding costs. A "Unified Fund" ID could not be chosen, so it was considered that all levels could be consolidated. Similarly, the two funds follow the same index, which is not the same investment product; the two accounts record the same fund and one of them should be deleted. It is the research data, the product entity or the record of the funds held that must be clearly stated before the real capital can be lost together. Such distinctions belong to the syntax of the field and not exclusively to Agent. Existing financial data services were maintained at that time, and direct reuse would be more reasonable than re-creating the centre of the Agent patent. The area framework should reduce the detail that must be understood by the caller, rather than copying an authoritative database.
- It's only a partial error.
Defines the amount as a currency decimal string that prevents CNY and USD from being inadvertently added; distinguishes the shares and net values into different amounts and helps check the unit relationship of the multiplier. However, the correct type does not support the correct identity of the fund or the net value derived from the applicable date.
Type systems, running time checks, source checks and business interpretations are therefore complementary. No one should be packaged as “financial correctness solved”. In particular, the adoption of a structured Schema should not result in credible certification of the entire report.
IV. Time is not an expired time, but a different set of issues
- Data describe the time with which we know it assumes that a teaching report describes the hold in September 30th, revealed on October 25th, Agent reads on November 2nd. The three dates were true, but they were different: the first describing the asset status, the second describing the time the information entered the public domain and the third describing the time the information was obtained by the system. “The latest report obtained today” cannot be summarized as “the holdout of today”. In turn, history does not lose value because it is not today; it can be used to explain the current configuration, to check past judgements or to construct studies with clear points of time. Effectiveness depends on the task, not the one-size-fits-all threshold. The report, which was only disclosed on 25 October, could not be considered to be known at that time if a judgement could be made on 1 October. This example shows that the deadline for the retention of data is not sufficient and that disclosure must sometimes be kept at the same time or the time is known to avoid taking subsequent information back to the past. Figure 2: Three times the same information
Three times, two research questions -- teaching date; the same information does not equal the same application to open the full vector Figure
- When new information comes up, do not quietly rewrite the old conclusions, assuming that users read the report and then import a corrected version of the warehouse. The new version should influence subsequent calculations, but the old report already publish had its own source and historical significance. Covering the underlying records directly and allowing old reports to automatically display new figures would prevent users from knowing what they were seeing at the time. The more rational design is to link the revision of facts to the revision of results: the old result still binds the old input and the new result binds the new input; the presentation explains the difference. The scope that needs to be recalculated is determined by real reliance, rather than discovering a change in one source and rerun all historical studies. Nor is it an order to keep everything permanently. The preservation strategy still requires compliance with permission, data minimization and application of retention rules. The way the version addresses the identification of different facts does not justify the indefinite retention of private information.
- Time strategy allows exit.
The source and version of the record may be sufficient for a one-time reading of the public statement without having to construct a two-time database. The semantic of additional time is worth entering into the performance contract only if the mission requires historical availability, cross-reference or revised tracking.
This principle is universal: the concept that can be distinguished when interpreted does not all immediately become the system field. First, it is determined which judgement, which calculation or which resumption of the act would change the distinction before deciding whether to increase the structure. Otherwise, the field modeling will move from solving real problems to preserving encyclopedias.
A denominator error sufficient to render “accurate penetration” meaningless
- Pass the teaching combination through the company X continues to use the previous combination. It is assumed that of the same calibre information identified, the weight of A on company X is 8 per cent, B 10 per cent and C 6 per cent. Company X ' s known exposure to total assets was 45 per cent x 8 per cent + 30 per cent + 10 per cent + 15 per cent + 6 per cent, or 7.5 per cent, corresponding to 9,000 yuan. It's just a teaching calculation. It was founded on the premise that the Fund ' s weight and the bottom weight were defined in a consistent manner, that the input applied to the time stated, that the company ' s identity was correctly mapped and that there were no unprocessed layers of duplication. The formula is simple and does not mean that it is simple. It is assumed that 60 per cent, 70 per cent and 40 per cent of the bottom disclosure coverage can be obtained by A, B and C. The Fund ' s bottom coverage was disclosed at 54 per cent, weighted by total assets. Cash also accounts for 10 per cent and the remaining 36 per cent of total assets have not yet been accounted for by this information.
The company X accounts for 7.5 per cent of total assets; it accounts for 7.5 per cent of the Fund ' s assets and 90 per cent of the Fund ' s assets, or 8.33 per cent; it represents 7.5 per cent of the Fund ' s sub-total sample, or 54 per cent, or 13.89 per cent. All three calculations can be correct, but answer different questions. If the system places the third number in the column “Combining actual company weights”, there is a misapplication of sample. The problem is not the decimal precision, but the structure known as sample is treated as the structure of the whole combination. Adding more decimal places will only make mistakes more authoritative. The missing information cannot be filled in to zero. Zero means we know no exposure, and unknown means we don't know if there's any exposure. Only when the assumptions of non-negative, non-leveraging, etc. are clearly met, can known portions be interpreted as corresponding downwards; this teaching assumption cannot be extended to all funds containing derivatives, empty heads or net openings. Figure 3: Same value, three different denominators
Same amount, three denominators - Teaching example: company X known amount 9,000 dollars open full vector Figure
- Why do cash targets have to use the adjusted denominator?
The original cash balance was $12,000. After the addition of 20,000 yuan, the total assets became 140,000 yuan without taking into account the teaching assumptions of market changes, costs and other cash flows. The retention of 10 per cent of cash means that the final cash should be C$ 14,000, so the funds available for the new funds are C$ 18,000, not all C$ 20,000.
If the system still calculates the target cash at the original level of $120,000, the full reserve of $12,000 will be used. Such errors do not require complex financial knowledge, but rather a distinction between “current combination” and “adjusted combination”. However, in multiple rounds of dialogue, new amounts, target ratios and old snapshots are easily mixed.
This is why the field approach is worth covering: The same cash continuity and denominator rules may be repeated for consultations, rebalancing, input comparisons and weekly reports. It is difficult to ensure that these portals remain in the same sense for the long term, by allowing each model to generate a calculation step on a provisional basis.
Determination calculations may also use the wrong method
- The correct implementation of the formula does not mean that it applies to the problem. In practice, the code can accurately execute an error formula or calculate the wrong data object. We need to distinguish between arithmetical correctness, the application of methods, the entry of truth and reasonable conclusions. For example, the net value of a teaching fund unit has changed from 1.00 to 0.99, with each cash allocation of 0.04 over the period, assuming that the allocation is no longer invested, that there are no other cash flows and cost adjustments. The total return held for a single period was 3 per cent, not negative 1 per cent of the net change. If the tool only calculates the net difference, even using unlimited precision is the wrong answer. If reinvestments are allocated, the date, price and share of reinvestment will also be specified; where external financial flows exist, the return measure will be selected to match the problem. It is not necessary to list all the financial formulas for the articles, but only to understand: the name of the algorithm, the calibration and the applicable conditions are part of the outcome.
- Costs need to distinguish between assumptions, estimates of terms and actual occurrence “Assumption rate is a value” “Estimation of costs based on currently applicable product terms” and “Account confirmation shows actual deductions” are three different facts. The first could be used for sensitivity analysis, the second required specific shares, channels, holding periods and validity dates, and the third required genuine evidence of enforcement. They should not share an unspecified “fee” field. In particular, where certain costs are already reflected in net values, a deduction from net proceeds may result in double-counting; the need for adjustments should be determined by a clear methodological contract, rather than automatically by the model seeing the word “rate”. Financial rules will change and product provisions will be revised. Therefore, the so-called domains are not variables that “a certain rate is fixed forever”, but rather “any cost estimate must know the applicable version of the article and the basis of the calculation”. Stability is the relationship necessary for judgement, not necessarily the parameter of the relationship.
- Don't set up another financial centre for unification.
When the different modules already have a clear methodology, the harmonization focus should be on a common semantic interface with results, rather than moving all formulas to a global Manager. Portfolio cash, attribution of proceeds, cost estimates and bottom exposure can be the responsibility of the most appropriate maintainer, respectively.
It may be worthwhile to share unit calculations, decimal rules, input references and version engagements; differences should be maintained between different revenue windows, cost sequencing and statistical methods. A formula that really repeats needs to be consolidated, with different meanings, but with similar names.
This federal design allows capacity to mature gradually. When there is only one consumer, correct errors are made in the original module; multiple missions encounter the same problem repeatedly before extracting deeper public interfaces. The abstract should reduce the knowledge of the caller rather than replace the original complexity with a longer configuration.
VII. The chain of evidence should bind claims, but not pretend to be a real machine
- Distinguishing sources, calculating and interpreting at least one line of data in the Fund's report is a source fact, and the concentration calculated by the process on the basis of the data is a derivative result, according to which the model explains that the “assembly is more risk-sensitive” is a study judgement. All three are important, but they have different evidentiary identities. Nor is the summary prepared by the sub-agent automatically a new original source. The summary may assist in locating the main agent, but when the principal agent has to express a critical amount or quote a specific table, it should be able to return to the corresponding version of the raw material or definitive result. An expert cannot be assigned the same status of fact as an official disclosure by simply assigning an ID to the expert. The establishment of such a relationship does not require each sentence to be accompanied by a huge object of evidence. The key is to retain the smallest link that can underpin important conclusions: which information, which location, what input, what method, which outcome, and which part of the conclusion is actually used. Figure 4: Evidentiary support claims, non-authentication claims
Evidence supports, does not authenticate automatically - source, computing and research advocates keep different identities open full vector Figure
- There's a link, Hash and blood, and there's still the wrong answer. A link can point to a real page, but it doesn't support that line. A Hashi can prove that the retained byte did not change, but that the byte was from the right product. Data blood is a complete record of the calculation path, but still a complete record of the wrong method. Therefore, the evidence system first provides testability rather than assurance of correctness. It makes it possible to ask, recalculate and locate questions and to reduce opportunities for errors to be hidden in natural languages. The value of such work is more accurate than the assertion that “the whole chain is credible and therefore the research is reliable”. Independent review also requires conditions of independence. If the second model reads only summary as written in the first model, it may simply repeat the same error. More meaningful reviews should, if necessary, be counter-argued on original grounds, stand-alone calculations and in different directions, rather than scoring fluidity.
- Maintaining partial conclusions is more useful than pursuing global credible booleans
In the portfolio cases, the duplicates of the A and B benchmarks confirm that the complete bottom exposure of C is not confirmed and that the cash algorithm of the additional funds can be independently established. These conclusions should be kept within each other's boundaries, rather than being crushed into a “report credible/untrustable” total switch.
The actual output allows a clear distinction to be drawn between “confirmed” “situations based on a dominant assumption” and “information still lacking”. When the necessary input is insufficient, the calculation should not result in a false result; however, the optional additional examination fails and independent facts should not be automatically set aside. Partial failure and overall delivery should be treated as genuine reliance.
It is not a relaxation of financial quality, but rather the avoidance of quality governance as a system of excessive disruption. What the users really need is to know which judgements are available and which need to be backed up, rather than seeing a red label that appears strict and does not actually explain why.
VIII. Autonomy must be in the research strategy, not in the mandate.
- The model can determine what next step is looking at. When it is found that C holds incomplete, the model can find the Fund Manager's report, compare the different disclosure windows, or judge that only a strategic risk interpretation can be made. It could also raise a more appropriate question: first, to identify duplication with the benchmarks and then to discuss whether additional funds could reduce the known concentration. This strategy adjustment is the value of Agent. If the system is to kill “first check A, then check B, then check C, and must produce an accurate overlap rate”, even if the model has found that the questions cannot be answered by their original calibre, it can only continue to complete an error plan. But the flexibility of research strategies does not mean that all rules are determined by models. Current user identity, readable accounts, suppliers allowed to send information, investment books and permission awards already awarded should not be altered by the model's perception that “it is easier to do so”.
- Research, recommendations and implementation need to be differentiated, but not always approved. A read-only combination analysis should also not automatically start a full set of transactions as a result of a “reduced concentration”. Control should match actual actions and consequences, not be tied to a financial keyword. When the task involves the private information of the client, data access and use is checked; when the mission requests that the true account be written, the authorization and conditions of validity of the specific action are checked. What is discussed here is the separation of engineering responsibilities and not the formulation of legal applicable conclusions for any region, institution or service type. “Users say yes” must also correspond to specific objects and versions. The authorization to view a particular hold is not the same as the approval to send to all external models; the approval of the operation of a certain amount is not the same as the continuation of the subsequent change. The binding of authorizations and actions is more important than approving bullet windows. Figure 5: Exploration can be open and the actual effects require a clear mandate
Open exploration, clear execution of authorization - research autonomy is not equivalent to external writing on permission to open full vector Figure
Open exploration, clear execution of authorization - research autonomy is not equivalent to external writing on permission to open full vector Figure
- Prompt is a way to explain, not a secure border.
It would be helpful to tell the model “read only” “do not trade” “do not divulge accounts”, but if the bottom tool still provides any database for writing or unrestricted access, the true permission boundary is not established. When the model deviates from the instruction, the system should still be able to limit the effects it produces.
On the other hand, all judgements should not be transposed to mechanical rules. For example, “is this study convincing” cannot be replaced by a fixed keyword. The certainty procedure deals with identifiable authorizations, identities and arithmetics; semantic quality is handled through appropriate assessment and review, and the two cannot pretend to each other.
IX. Protracted preservation of the state of implementation, with additional implications for continuity of research
- Restoring a session does not mean resuming a study of the problem. LangGraph clearly distinguishes between checkpoint and cross-line storage; Claude Agent SDK also supports continuation of the session. These capabilities are worthy of reuse rather than being treated as non-existent by proxies. [2][4] The additional question for financial research is: Are the same questions still answered at the time of recovery? What period? Has the user changed the target? Should the phase conclusions of publish be retained or amended? Is the original data authorization still valid? These implications need to be clearly expressed in terms of application or field logic. A session record can preserve all the conversations, but it remains impossible to distinguish between “the assumptions of last week” and “the facts of this week”. Thus, long-term research is not just a longer history of chat, but rather the ability to relate objectives, rationale, conclusions and reasons for change.
- Don't read it blindly and repeatedly. The latter may have produced results, and simple retry may lead to duplication. Even if a report is saved, a distinction should be made between failure before writing, successful writing but confirmed loss, and a state that is not yet known. LangGraph's aborted document is a clear reminder that the node code before the interruption occurs may be re-executed and therefore the associated side effects need to be designed, for example, with a sauerkraut. The problem is not a flaw in a given framework, but rather the application of operative semantics that need to be shared with the operation. [5] The contribution of financial niche systems should be to provide identification, retry and cross-checking for specific business operations, rather than claiming a common “just one time” switch. Where there is no corresponding protocol for the external system, the maintenance of a successful mark in the local database does not create an off-site assurance.
- A lot of missions don't need long-term memory.
A one-time comparison of the two publicly available materials without a long-term preference and without the construction of a cross-week study area may be the best design. long-term memory, which would lead to obsolescence, conflict, permission and delete semantics, should be driven by real and continuous tasks.
Where follow-up is necessary, priority is also given to preserving valuable stabilization preferences and research objectives, rather than depositing each input generated into reality. In the past, the model stated that “the fund is only for you” and that it cannot be upgraded unconditionally into a long-term investment preference already identified by the user in a follow-up session.
X. An economic boundary exists between area capacity, Skill and the framework
- If a task requires a set of research steps, such as identifying the fund, looking for a report and choosing the applicable indicator, then Skyll is the appropriate expression. It allows the model to load according to task, without having to place the entire field of knowledge into a resident system alert. If the key to the task is a repeated certainty formula, the formula is placed in a tool or field module. The parameters, units and failure conditions required for the tool should be clear, and Skill is responsible for indicating when it will be used. Don't use a longer Skill instead of a rule of calculation that can already be programmed. Area SDK or framework begins when multiple tasks share the identity of the target, the time-point calibre, the source reference and the result structure, and these rules need to be independently evolved and validated. The framework is the result of a reuse demand and should not be a directory template that must be filled in at the start of the project.
- The value of a frame, which can be tested by deleting experiments, envisages the deletion of a public module. If the complexity also disappears together, the caller will understand more easily, and it may be merely a worthless repeater layer. If, after deletion, the same cost calibre and source verification is redispersed to multiple entry points, the module does manage the complexity centrally. This experiment is more meaningful than "Does the code have a layer of interface?" The depth of the interface derives from how much detail it hides that the callor should not have borne, rather than how abstract the type name is. An abstraction with only one adaptor may be useful, but it should be acknowledged that portability at the time of operation has not yet been physically validated. The same applies to independent deployment. Logical responsibilities can be separated but need not be broken down into micro-services immediately. Cross-process costs can only be affordable if deployment isolation, resource mobilization or organizational collaboration generate real needs.
- The normal mission is a complex architecture paid.
Does the additional mechanism add context, assembly steps, database access and failure face when it is not triggered? Is there a need for a new tool to simultaneously modify multiple lists, restore protocols and tests? These are the normal path costs of the framework and the incremental costs of change.
This does not create a uniform percentage for them. The different deployments and tasks are very different and should be measured by real requests. However, costs cannot be assumed to be non-existent without quantization, much less simply demonstrating the benefits of complex mechanisms in extreme malfunctions, overlooking the additional daily burden of ordinary inquiries.
Thus, a reasonable extension should normally be activated as needed: short missions do not start a long-term workspace, no independent sub-issue does not start an expert, the results are available when sufficient, and the option is not to delay an independent legal answer when unavailable.
XI. The strongest objection: is it not enough to write the rules directly into the existing framework?
- Yes, and often the right answer, with Claude Agent SDK running research, MCP exposure fund tools, plus an area calculator, can be a complete available option. The same is true for reading, computing and manual checking using the LangGraph organization, or extending access to the same tools with Pi. As long as the rules are centralized, the tests are clear and the results can be verified, there is no requirement that a package called FinancialAgentFramework should appear in the code warehouse. There is no natural rating of value between the stand-alone framework and the logical business module, and the determining factor is actual reuse and maintenance effectiveness. Therefore, this paper does not attempt to prove that "all financial actors need to study their own framework." On the contrary: the complexity of the field proves the need to model the field and does not automatically prove the need for a new running time. Those who advocate the addition of an implementation system need to supplement the rationale for independence at the executive level.
- Where it is possible to support a self-build operation where a key performance contract cannot be satisfied under actual conditions of deployment when it is already in operation, where the extension points are insufficient and the compatibility layer is more complex than a partial self-study and deserves careful comparison. Possible conditions include special model agreements, access to resources, inter-process recovery or irreplaceable delivery synonyms, each of which requires specific back-up support. “We want to be better controlled” is not specific enough. Clarification should be given as to which actions, which failures, how the existing framework was being addressed, why it did not meet operational requirements, and whether a smaller expansion had been attempted. Otherwise, the so-called control power simply moved the maintenance of mature frameworks to its own team. Even for eventual self-establishment, replacement costs should be factored in: protocol flow fractions, elimination of competition, retry, plugin life cycle and state migration can consume work time over time. These may be of research value, but cannot automatically be recorded as differences in financial research. Figure 6: Proceeding from the smallest options rather than the platform
Reuse first, then prove that incremental is necessary - the field core is two kinds of judgment when running. Figure
Reuse first, then prove that incremental is necessary - the field core is two kinds of judgment when running. Figure
- A reverse success standard.
Migration to a more mature generic framework that removes a large number of self-owned codes of execution while keeping the Fund ' s calculations, evidence and permission behaviour unchanged should be seen as success rather than a technical retreat. The truly accumulated area capacity should not depend on maintaining the size of self-study codes.
On the contrary, if the model is changed each time, it is necessary to rewrite the Fund to indicate that the semantics of the field are too closely aligned with the details of implementation. The value of the pristine core lies in stabilizing reusable judgmental relationships, not in the need to bind a generation forever to Harness.
XII. The necessity of Agent also requires separate proof
- Multiple roles do not amount to multiple independent messages, leaving the same information to the “multi-head” “scientist” “risk expert”, which may be worded differently and may be counterproductive, but it does not guarantee the true independence of the information. All three models were based on the same error summary, and the final vote would only add to the wrong momentum. The clearest engineering value of sub-agents is usually the isolation of context, the decomposition of parallel sub-issues, and the use of specific tools and methods. The official Claude Agent SDK also listed these as important uses of sub-agents. Improved financial judgement also needs to be measured in specific tasks. For example, one expert is responsible for checking the point of disclosure, another is responsible for the recovery of cash and costs, and the third is looking for evidence to reverse the configuration assumptions, which is closer to a useful division of labour than simply naming three individuals. However, the division of labour may not have a net gain if the principal agent has been able to complete these checks at low cost.
- The main agent cannot simply receive “experts say no” a valuable handover should include what the expert answers to the sub-questions, what information is based on, what methods are used to calculate, and what gaps remain. The agent needs to know which are the original facts and which are the expert opinions, and can return to key raw materials if necessary. Otherwise, we simply add a natural language agent layer between the main agent and the facts that cannot be checked. context, although shorter, has lost its responsibility for research. The greater the number of experts, the greater the complexity of the information permission, the budget, the cancellation and the recovery.
- Write the exit conditions in the design.
No experts are activated in the absence of an independent sub-issue; no more role is required to compensate for the lack of information; and when the evaluation does not reveal additional errors and significant delays, it returns to the single agent baseline. The exit mechanism does not lack confidence in the senior structure, but rather avoids transforming an organizational pattern into a fixed cost that must be paid for the long term.
The goal of Finance Agent is to better complete the study, not to show how much it can run at the same time. What has been improved by multiple agents requires a simple baseline comparison with the same subject, the same information window, comparable model conditions, and not just a lively and successful demonstration.
XIII. EVALUATION SYSTEM TO RESPONSE QUESTIONS OF DEPARTERS, NOT ONLY STATISTICAL INSTITUTIONS
- Implementation, research, delivery and adoption are different observations One end of operation, indicating that the implementer is in its final condition; success in the preservation of the report, indicating that certain bytes have been dropped; user capability to download, indicating that the delivery chain has been established; and independent review of the conclusions to indicate that the quality of the study is supported to the extent appropriate. These facts are interrelated and cannot be replaced by one another. Similarly, users understand and use a study and do not directly prove that it has increased investment returns. The investment results are influenced by factors such as markets and behaviour, which do not allow long-term returns from Project acceptance check. The separation of the different perspectives does not erase progress, but allows for an accurate evaluation of each. Figure 7: Four completed, not an automatic upgrade chain
Four completes, check each other -- the view is related, but not automatically upgraded to open the full vector. Figure
- The evaluation should contain counter-circumstances that can change conclusions. Change
Auto-fetched and translated by AI Atlas. All rights belong to the original author.