ETF Agent 工程实践：从个人研究方法到可恢复的 Subagent
如果希望用户创建自己的 AI 专家，最简单的实现方式是什么？
做一个设置页面，填写名称、描述和 Prompt，保存到数据库。下次用户选择这个专家时，把 Prompt 放进模型请求。
从演示来看，似乎已经足够。但当我们真正将自定义专家接入 ETF Agent 的研究系统时，问题远不止如此：用户编辑了专家，正在运行的研究要不要立即采用新版？模型怎样发现专家，又怎样避免把所有人的完整 Prompt 都塞进主上下文？研究中断后恢复，究竟该使用哪个版本？如果某位专家写着“访问全部数据”，它真的获得了这些权限吗？
这些问题决定了“自定义专家”是一段临时提示词，还是一项能够长期维护的产品能力。
我们选择不为每位专家创建一套 Runtime，而是让用户保存一份带有身份和版本的研究方法，交给既有的通用 Subagent 执行。实践中也踩过一些很具体的坑：指引进入了错误的请求层级，Child 明明使用了个人专家，History 却识别不出它，还有一次浏览器验证因两份 React 实例而出现空白页。
这篇文章只讨论用户自定义专家的保存、选用、执行和恢复。Web/PDF/Python 能力包与多 Agent 对话，是它所依赖的基础，而不是本文要重复介绍的系统全景。
一、为什么不能继续给 Child 叠加一个角色 Prompt？
ETF Agent 原本就有 Web、PDF、Python 子代理和六种内置投资方法。前者指向资料阅读与计算的工作侧重，后者代表 Markowitz、Taleb、Bernstein、Bogle、Marks、Fama-French 等不同研究框架。
这些方法由开发者维护。可用户自己的习惯未必符合固定目录：有人要求始终检查基金费率与跟踪差，有人先穿透底层持仓，也有人希望每个判断都附带证伪条件。把这些习惯写进系统 Prompt 会影响所有任务；让用户每次重述又无法积累方法。
早期接口分别用 type、perspectiveId 和 subagentProfileId 表达执行专长、投资视角与个人配置。这是按需求逐步增加的合理设计，但随着专家种类增加，调用者开始需要理解多个维度的组合。
我们没有简单新增一个 custom 类型，或只在界面隐藏复杂参数。统一普通 Child 的选择接口改成：
{
"agentId": "user/<profile-id>",
"task": "根据已提供的披露资料核对基金底层持仓重叠"
}
这里的 ID 仅作形状示例；实际 user/ 后面是服务端签发的 Profile UUID。不选专家时使用 builtin/general。一次普通 Child 只绑定一份专家定义；执行机制、模型路由和授权并不随专家名称重新创建。
用户保存的是“怎样研究”，而不是“另一个拥有独立权力的 Agent”。 这成为后续设计的分界线。
二、保存的是方法，不是运行平台
第一期个人专家故意保持很轻：
- name：用途：用户识别和管理时显示的名称。
- description：用途：Main 判断何时适合使用这位专家。
- instructions：用途：选中后交给 Child 的完整研究方法。
- enabled：用途：是否向后续新研究开放。
例如用户可以创建“基金持仓重叠检查员”。描述说明它核对多只基金的底层资产集中度；指引要求先检查持仓披露日期，区分产品名称与底层公司，资料不足时保留未知，而不是输出精确暴露的猜测。
这只是一个教学配置，不是 ETF Agent 对真实基金结果的认证。它说明个人专家最值得保存的内容是判断顺序、关注点和可重复使用的约束，不必是富有戏剧感的虚拟人设。
服务端为个人配置分配稳定的 profileId，保存 revision、内容哈希与账户归属。保存或更新时，客户端提交 expectedRevision，数据库通过条件更新防止并发覆盖。删除是软删除：停止新研究继续发现此专家，但不把已经发生的研究改成“从未使用过它”。
还有容量界限。当前配置限制是最多保存 20 个 Profile、启用 8 个；这些限额服务于账户存储与模型目录的可控性，不是每位专家的推理预算。
当保存失败、发生版本冲突或超过容量时，UI 需要保留用户未提交的编辑，不能把错误提示当作一次成功保存。受控浏览器测试中，两页同时编辑同一 Profile：先保存的新版生效，旧修订被拒绝且未提交草稿保留。这里使用的是合成身份和进程内存储，不能当作生产登录验收。
三、从设置页到研究结果，一条链如何走通？
图中关系（文字对照）：
- 账户设置 → 已保存背景。
- 已保存背景 → 运行快照。
- 运行快照 → 主代理目录。
- 主代理目录 → 选定专家。
- 选定专家 → 通用子代理。
- 通用子代理 → 结果回执。
- 结果回执 → 主代理回答。
- 结果回执 → 历史记录。
这张图是关键事实的交接关系，并不是一条必须按固定速度顺序执行的队列。
1. 新 Run 冻结自己的目录
研究开始时，系统读取该账户已经启用、未删除的个人专家，并将当时的内容与版本固定到这个 Run。
Main 收到的是有界的目录：精确 ID、名称、描述。它不需要一次性加载所有专家的完整 instructions。只有选中某位专家，完整方法才作为带有用户来源标记的数据进入对应 Child。
目录为空与目录读取失败是不同事实。后者需要反馈“当前专家配置不可用”；系统可以继续其他获准研究，却不能悄悄用 General 替代一位被明确要求的专家。
2. 由同一套 General Child 执行
Child 沿用既有模型、工具、来源授权、资源管理与结果交付。个人专家的说明并非新的系统指令，也不授予交易、发布、递归创建 Child 或额外数据访问权限。
这并非只靠 Prompt 写一句“不要越权”就能保证。在一次受控请求装配回归中，迁移到统一 agentId 后，个人 instructions 曾误入 Child 的通用 system 层。我们沿原请求装配纠正来源位置，确保 Main 只得到目录、选中的 Child 才收到冻结方法。一段指引是否可信，还取决于它来自谁、以什么角色进入最终请求。 这是工程测试发现的问题，并无真实用户数据泄露的证据。
一次执行选择的准确身份包含 agentId、来源、版本、内容哈希和个人修订号。agentId 不是文件路径，也不是执行脚本的许可证。
当前用户自定义 Profile 是文本型配置。与此相邻，内置 PDF 等受管理专家已支持 SKILL.md、参考文件和受控 Python 辅助脚本；但用户目前不能因此随意上传 ZIP 并执行任意代码。共用选择入口，不等于共用全部包安装能力。
3. Main 只依据实际结果交付
Child 完成工作后，原执行系统记录实际结果。Main 可以读取结果并综合回答；History 根据原 Child 身份和绑定版本显示实际使用的个人专家。
迁移期间，Child 已使用新的 agentId，旧 History 读者却因识别旧字段而漏掉专家归属。修正不是凭专家名称推测，也不是拿账户中的最新版补旧记录，而是回到原 Child 的准入、冻结方法和真实执行回执。这样，日后的改名、停用或软删除都不会改写当时使用的版本。
“创建成功”“被选中”“Child 已准入”“Child 已交付”是四件不同的事。系统不能因为界面显示了一个角色名称，就声称该专家真正完成了研究。
4. 旧研究不能偷用新版本
一个反例足以说明为什么需要 Run 快照。
上午使用 Profile 第一版启动基金研究。下午用户把它改为第二版，增加更严格的披露时点要求。晚上研究从暂停点恢复。
如果恢复时只凭 profileId 重新查最新配置，同一次研究前后就可能使用不同方法，却显示同一个名字。因此我们保持原 Run 的冻结内容：新编辑只影响后续新 Run，旧 Run 沿原绑定继续。
这既是版本一致性，也是解释研究结果时必要的事实基础。
四、对照 Octop：模板、实例和执行不必是一个东西
腾讯云开源的 Octop 展示了另一种专家产品路线。它的专家模板可以包含 manifest.json、SOUL.md、IDENTITY.md、Skills 等文件，实例化时把模板内容写入独立 Agent 工作区。用户随后持续使用、修改，还可以通过受支持的机制分享或发布模板。具体模板复制与目录机制可参见其 Expert Catalog 源码。
Octop 更适合“长期独立工作的数字助手”这个产品问题，而 ETF Agent 当前的个人 Profile 主要回答：“一次完整研究怎样准确采用用户保存的某套方法？”
两条路线各有合理性。独立专家实例便于保留自己的对话、文件和工作习惯，但需要管理更多长期状态。我们则让个人方法进入原 Run/Child 体系，不复制模型、记忆、预算和恢复栈；代价是个人专家暂时没有独立的长期对话空间。
比较之后，我们更确信应该分清三个对象：
- Expert Definition：这位专家采用什么方法，哪个内容版本；
- Expert Execution：这次在什么授权范围内实际做了什么；
- Expert Workspace：如果确有长期使用需求，才为它持有哪些持久工作资料。
未来可以借鉴 Octop 的专家卡片、示例任务、模板分享等体验，但不必仅为了一个 Prompt 配置页面，就创建第二套 Agent 平台。尤其在金融场景，模板共享不能自动带走个人持仓、私人研究历史或原授权。
五、用户真正怎么使用专家：从建议到真实贡献
原先的问题是“我能不能保存一个专家”；进一步实现后，问题变成“我在聊天里选了它，它到底有没有参与这次研究”。
我们没有把目录卡片设计成新的执行入口。账户可见的专家目录只展示身份、用途与示例，不返回私有完整指引；输入框里的卡片选择和明确的 @ 选择使用同一条精确 agentId 链路。用户原文保留，发送后由原消息与 Entry 保存选择，再由 Main 在合法条件下委派 Child。文本里偶然出现一个专家名字，不能凭字符串匹配变成一次授权调用。
这里有四层不能合并的事实：
- 专家目录、方法说明：真实含义：当前账户可以发现的配置；不能冒充：已获得工具权限。
- 显式选择、常用组合：真实含义：对本次研究的选择或建议；不能冒充：Child 已经派发。
- Child 请求、阶段回执：真实含义：原执行者实际开展了哪些工作；不能冒充：Main 已采纳该观点。
- Main 正文与 History 贡献区：真实含义：已提交的答案及可见执行依据；不能冒充：专家观点已达成共识。
例如，一个合成的基金研究问题需要核对 PDF 中的费率，并计算组合风险。用户选择 PDF 专家，另从常用组合取得 Python 专家的可编辑建议。Main 可以只派发部分合适的 Child，不能因为组合里有两位就显示“两位专家均已参加”。运行中 PDF Child 交出带引用的结果后，统一贡献区沿原 taskRef 和实际结果呈现它；Main 是否采用以及最终怎样解释，仍由 Main 的正文说明。
账户常用组合的 presetId、版本、名称和专家 ID 列表由原账户存储维护。应用组合只是填入可编辑建议，不自动开启团队房间、Peer 对话、Reviewer 或额外预算。刷新、历史重开只读原 Run 冻结的选择与实际贡献，不能用今天的目录为昨天的运行补一个虚构成员。贡献展示局部失败时，原正文与下载仍应独立可读。
六、让专家改变研究方向，而不改写历史
研究描述也不能被 Main 最初的一句话永久锁死。假设 Main 交给某位 Child：“计算这两只基金底层持仓的精确重叠”。Child 阅读披露后发现其中一只只有前十大持仓；继续执行原命令，容易把部分覆盖当作完整重叠。
我们把方法身份的冻结和本次研究任务的修订分开。专家包、账户修订号以及原 Run 身份保持不变；Child 则可在原 submit_subagent_result 的可选 update 中修订 task、hint、relationHint 和合法 inputRefs。已接受的修改只进入后续请求，不覆盖原创建记录、已派发请求或用户原话。
下面是抽象教学示意，不是原始模型输出：
{
"summary": "第二只基金仅披露前十大，无法计算完整持仓重叠",
"update": {
"task": "计算已披露部分的重叠，并明确覆盖率与不可推断部分",
"relationHint": "供主研究参考，不能将部分覆盖写为完整敞口",
"inputRefs": ["本次运行已授权的真实结果引用"]
}
}
新增输入仍经过原授权和来源检查；越界输入只能被局部拒绝，不能顺便取消已经合法的描述修订或清空独立正文。模型自己修改研究方法不意味着可以换一个专家身份、追加资源额度、放宽账户限制或把文字关系变成自动等待的任务图。尤其是 relationHint：它表达希望怎样协作，不代表另一位 Child 已执行、已发送消息或其结果已被读取。
这项演进与 OpenAI Agents SDK 对 agents-as-tools 与 handoff 的区分有相邻问题：这里选择让 Main 继续负责用户回答，而不是把对话控制权永久交给某个专家。但具体的版本化配置、软偏好、回执修订和账户展示，仍要由应用的原负责者实现。新增自由度值得投入的条件，是它能保留真实依据而不是制造更多模型自述。
七、这套机制已经走到哪里？
截至 2026 年 10 月 27 日，个人 Profile 的版本冻结、普通消息精确选择、专家软偏好与预加载、按需调用指南、目录/贡献展示、常用组合及 Child 自主描述修订，都已有真实实现与有界工程验收。隔离的 PostgreSQL、认证浏览器和脚本化模型请求用于证明选择、存储、回读、恢复与界面路径，不能代替真实模型是否会选对专家或采用偏好的质量评估。
更早的真实模型研究样本主要是内置研究方法或 General/Python，仍不能推断个人自定义专家已经在真实投资问题上提高质量。新增体验也不意味着允许账户随意安装可执行专家包、给某位专家一套独立预算，或者让建议文本获得交易授权。
这不妨碍解释一个已经成立的工程判断：个人研究方法可以作为版本化的数据，复用现有 Child 执行与结果系统；没有必要因此新增一个拥有全部权限的 Expert Manager。
结语
实现自定义专家最简单的一步，是提供 Prompt 输入框。让它真正可维护的，是方法身份固定、任务描述可修、实际贡献可重读，而不让“用户喜欢什么”变成“系统已经做了什么”。
ETF Agent 的做法不要求每新增一种个人研究方法，就额外维护一个模型实例、一份预算、一套授权或一份任务账本。用户可以逐步积累自己的研究习惯；系统则在每次执行中记录：用了谁的方法、采用哪一版、实际上完成了什么。
专家值得被复用，执行事实更值得被保留。
只有两者分离，用户定义的专家才能从一次性的角色扮演，演变成一种可以持续升级、又不会拖垮已有 Runtime 的能力。
延伸阅读与交流
Subagent：从上下文隔离到专业化协作 · 模型会理解，代码才授权 · ETF Agent 技术文章仓库（欢迎 Star）。
原文与交流
阅读原文 · ETF Agent 工程文集（欢迎 Star）。
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。
ETF Engineering Practice: From Personal Research Methodology to Recoverable Subagent
If you want the user to create your own AI expert, what is the simplest way to achieve it?
Make a settings page, fill in the name, description and Prompt, and save it to the database. Next time the user selects this expert, put Prompt in the model request.
From the demonstration, it seems enough. But when we actually connect custom experts to the ETF Agent research system, the question is much more: do users edit experts, do they want new versions of the research being run? How does the model find experts, and how does it avoid putting the entire Prompt in the main context? What version should be used to recover after the study has been interrupted? If an expert wrote “access all data”, did he really get these permissions?
These questions determine whether the “self-defined expert” is an interim prompt or a product capacity that can be maintained over time.
We choose not to create a set of Runtme for each expert, but rather to allow the user to save a study with an identity and a version, to be executed by the established generic Subagent. In practice, there have also been some very specific pits: guided to the wrong level of request, while Child used a personal expert, History did not recognize it, and a browser validated a blank page due to two React examples.
This article only discusses the preservation, selection, execution and recovery of user-defined experts. The Web/PDF/Python capacity package and multiAgent dialogue are the basis on which it is based, not the system landscape that this paper repeats.
I. Why can't we continue adding a role to Child?
ETF Agent originally had Web, PDF, Python sub-agents and six built-in investment methods. The former points to a focus on reading and computing, while the latter represents different research frameworks such as Markowitz, Taleb, Bernstein, Bogle, Marks, Fama-French.
These methods are maintained by developers. Users ' own habits may not necessarily be consistent with a fixed catalogue: there are requests to check consistently the difference between the Fund ' s rates and their tracking, and there are those who go through the bottom holder first, and there are those who would like to see perjury attached to each judgement. Write these habits into the system Prompt will affect all tasks; each time the user repeats it, the method cannot accumulate.
The early interfaces express executive expertise, investment perspectives and individual configurations using type, perspective Id and subagentProfileId, respectively. This is a rational design that increases in demand, but as the pool of experts increases, the caller begins to understand a combination of multiple dimensions.
We do not simply add a custom type, or hide complex parameters only in interfaces. Align common Child selection interface with:
{
"agentId": "user/<profile-id>",
"Task": "Reconcile the overlap of the Fund's base hold on the basis of the disclosure information provided."
}
The ID here is only an example of shape; the actual user/ is followed by the service-issued Profile UUID. Do not select an expert using the biltin/general. General Child bound only one expert definition at a time; implementation mechanisms, model paths and authorizations were not recreated with the name of the expert.
The user saves “how to study” rather than “another person with independent powers”. This is the dividing line for subsequent design.
II. METHODOLOGY, NOT THE ROLE OF OPERATION
The first group of individual experts was deliberately kept light:
- Name: Use: Name displayed when user identifies and manages.
- Discription: Usage: Main judge when it is appropriate to use the expert.
- uses: a complete study method selected and forwarded to Child.
- Enabled: Use: Open to new follow-up studies.
For example, users could create “overlay inspectors for funds”. The description indicates that it reconciles the lower asset concentration of multiple funds; the guidance requires that the date of the warehouse disclosure be checked first, distinguishing between the product name and the lower company, and that the information be kept unknown when it is insufficient, rather than output an accurate exposure guess.
It's just a teaching configuration, not an ETF Agent certification of real fund results. It shows that the individual expert ' s most valuable elements are the order of judgement, the focus and the constraints of reusable, and need not be a dramatic virtual person.
The service end allocates a stable profile Id, saves access, content Hash and account attribution. When saving or updating, the client submits an excepted review, and the database is updated to prevent and distribute coverage. The deletion was soft: the new study would stop continuing to identify the expert but not replace the study that had already taken place with “it has never been used”.
There are capacity limits. The current configuration limits are up to 20 Profiles and enabled 8; they serve the controllability of account storage and model directories, not the reasoning budget of each expert.
When saving failed, a version conflict occurred or a volume exceeded, UI needed to keep the editor that the user had not submitted, and the error hint could not be saved as a success. In the controlled browser test, two pages are being edited simultaneously for the same Profile: a new version saved first becomes effective, an old amendment is rejected and a draft is not submitted for retention. It uses synthetic identity and process storage and cannot be used as a production login to acceptance check.
III. How does a chain go from page setting to research findings?
Relationships in figures (text contrasts):
- Account Settings → saved background.
- Saved background runshot.
- Run a snapshot of the master proxy directory.
- The master proxy directory selects experts.
- Selected experts - General sub-agents.
- General Sub-Agent.
- The result is a reply from the agent.
- Turns out it's history.
This chart is the intersection of key facts and is not a queue that must be carried out in a fixed-speed sequence.
New Run Freezing Your Directory
At the beginning of the study, the system was able to read the individual expert that the account had been activated and had not been deleted, and fixed the content and version to this Run.
Main received a directory of boundaries: precise ID, name, description. It does not require the complete loading of all experts at once. Only if an expert is selected will the complete method enter the corresponding Child as data marked with the user source.
It is not a fact that directories are empty and directories fail to read. The latter required feedback that “the current expert configuration was not available”; the system could continue with other approved studies, but could not quietly replace a specifically requested expert with General.
By the same set of General Child
Child follows existing models, tools, source authorizations, resource management and results delivery. The individual expert ' s description is not a new system directive and does not grant transaction, publish, Creativity Child or additional data access to permission.
This is not guaranteed by Prompt's "Don't overstep the power". In a controlled request for assembly return, the individual moved to the universal system layer of Child by mistake. We follow the original request to assemble to correct the source and ensure that Main receives only the directory, the selected Child, to receive the freezing method. The credibility of a guide also depends on who it comes from and what role it takes to enter the final request. This was a problem discovered by the engineering tests and there was no evidence of a real user data leak.
The exact identity of the one-time implementation selection consists of angentId, source, version, content Hashi and individual revision numbers. Agent Id is not a file path, nor is it a licence to execute scripts.
Current user-definedProfile is a text-type configuration. Next to this, managed experts such as the built-in PDF have supported SKILL.md, reference files and controlled Python-assisted scripts; however, users cannot currently upload ZIP and enforce any code at their disposal. Shared selection of access points is not equivalent to sharing full package installation capacity.
Main delivery based on actual results only
After Child completed its work, the original implementation system recorded the actual results. Main can read the results and provide a comprehensive answer; History shows the actual use of individual experts according to the original Child identity and bound version.
During the migration, Child has used the new agent Id, and old History readers have missed expert affiliation by identifying the old field. The amendment was not based on the expert name, nor on the latest version of the account to replace the old record, but rather on the original Child access, the freezing method and the genuine execution response. As a result, no later change of name, discontinuation or soft deletion would be rewrite the version used at that time.
“Create succeed” “Child has access” “Child delivered” is four different things. The system cannot claim that the expert actually completed the study because the interface shows a role name.
4. Old studies cannot steal new versions
A reverse example is enough to explain why Runshot is needed.
In the morning, the first edition of the Fund's research was launched using Profile. In the afternoon, users changed it to the second edition, adding stricter disclosure point requirements. At night, research was conducted to recover from the pause.
If only profileId is used to recheck the latest configuration, different methods may be used before and after the same study, but the same name is displayed. So we keep the original Run freeze: the new editor only affects the subsequent new Run, the old Run continues with the original binding.
This is both a consistent version and a necessary factual basis for interpreting the findings of the study.
IV. Optop: Templates, examples and implementation need not be an object
Tencent cloud open source Octop shows another line of expert products. Its expert template can contain documents such as documents such as minest.json, SOUL.md, IDENTITY.md, Skills, etc., and will be translated into an independent Agent workspace. Users continue to use, modify and share or publish templates through supported mechanisms. A specific template copy and directory mechanism can be found in its Express Catalog source code.
Octop is better suited to the product question of a long-term stand-alone digital assistant, while the current personal EF Agent Profile answers, inter alia: "How exactly does a complete study use a user-saved method?"
Both routes are reasonable. The independent expert ' s example facilitated the retention of his dialogue, documentation and working habits, but needed to manage more protracted situations. We allow the individual approach to the original Run/Cild system, without replicating models, memories, budgets and restoration; the cost is that individual experts are temporarily without an independent and permanent space for dialogue.
After comparison, we are more convinced that three objects should be distinguished:
- Expert Defense: What is the expert's methodology and which content version;
- Expert Exchange: What actually was done this time within what mandate;
- Express Workspace: What durable work information is available for it if there is a need for long-term use.
In the future, you can draw on Optop's expert cards, illustrative tasks, template sharing, etc., but without having to create a Prompt configuration page to create the second Agent platform. Especially in the financial arena, template sharing does not automatically take away individual holdouts, private research history or original authorization.
V. How users really use experts: from recommendations to real contributions
The original question was “Can I keep an expert”; when it was further realized, the question was “I chose it in a conversation, whether or not it was involved in the study”.
We didn't design the catalogue card as a new implementation portal. The expert directories visible to the accounts show only identities, uses and examples and do not return to the private complete guide; enter the card selection in the box and specify @ choose to use the same exact entry Id link. Keeps the original user, sends after the original message and the Entry saves the selection, and Main assigns Child under lawful conditions. The occasional name of an expert appears in the text and cannot be converted to an authorized call by string matching.
There are four layers of facts that cannot be merged:
- Directory of experts, description of methods: true meaning: configurations that can be found in the current account; not impersonating: tools permission have been acquired.
- Visible selection, common combination: true meaning: choice or recommendation for the study; no impersonation: Child has been sent.
- Child ' s request, stage reply: the true meaning: what was actually done by the original implementer; it cannot be assumed that Main has taken that view.
- Main Text and History Contribution Area: The Real Meaning: The Answers Submitted and the Visible Basis of Implementation; It cannot be assumed that the expert opinion has been agreed.
For example, a synthetic fund study would need to reconcile the PDF rates and calculate the portfolio risk. User selects PDF experts, and obtains editable suggestions from Python experts from common combinations. Main can send only the right part of Child, and it cannot be shown that “both experts have participated” because there are two in the combination. The unified contribution area presents the results of the running PDF Child after handing over the quoted results, along with the original taskref and the actual results; whether Main is used and how it is eventually explained is still explained in Main's body.
A list of presetId, version, name and expert IDs of the account's frequently used combination is maintained by the original account storage. The application combination is only a editable proposal and does not automatically open the team room, the Peer Dialogue, the Reviewer or the extra budget. To refresh, re-read and re-read the Run freeze options and actual contributions, today ' s catalogue cannot be used to fill yesterday ' s run with a fake make-up. When the contribution shows partial failure, the original text and downloads should remain independently readable.
VI. Redirecting of experts rather than rewriting history
The research description cannot be permanently locked down by the original line of Main. Assumptions that Main hand over to someone Child: “Accounting the precise overlap between the two Fund's bottom warehouses”. Child, reading the disclosures, found that one of them had only the top 10 holding tanks; continuing the original order would make it easier to treat partial coverage as a complete overlap.
We separate the freezing of methodological identities from the revision of this study mandate. The expert bag, the account revision number and the original Run identity remain unchanged; Child can amend the name of the user, hint, relationHint and legal inputRefs in the original option for submit subagent result. The accepted changes go only to follow-up requests and do not cover original creation records, sent requests or user words.
The following is an abstract pedagogy, not an original model output:
{
"Summary": "The second fund only discloses the top 10, and cannot calculate the full overlap."
"update": {
"task": "calculating the overlap of the disclosed part and specifying the coverage and the non-in extrapolational part,"
"relationHint": "The main research reference cannot be written in part as full open".
"input Refs": ["The Real Results of this Operation"]
}
}
The new entry is still subject to the original authorization and source check; cross-border entry can only be partially refused, and the already legal version of the description can not be cancelled or the independent text cleared. The model does not imply a change of expert status, additional resources for quota, the relaxation of account restrictions or the translation of text relationships into automatic waiting task maps. In particular, relationHint: It expresses its desire to collaborate, not on behalf of another Child who has been executed, sent or whose results have been read.
This evolution has an adjoining problem with Open AI Agencies SDK regarding the distinction between the sexes and handoff: It was chosen that Main would continue to be responsible for the user's response, rather than leaving control of the dialogue permanently to an expert. However, specific version configurations, soft preferences, buy-back revisions and account displays are still to be achieved by the original operator of the application. The condition that the new freedom is worthy of input is that it preserves the true basis and does not create more models.
Where has this mechanism gone?
As of October 27, 2026, the individual Profile version was frozen, the generic message precision selection, the expert soft preference and preloading, the call-on-demand guide, the catalogue/contribution presentation, the common combination and the Child self-presentation revision, all of which have been realized in real terms. Separated PostgreSQLs, authentication browsers and script-based model requests are used to demonstrate whether the selection, storage, rereading, restoration and interface paths will be selected as experts or whether a preferred quality assessment will be used for a real model.
An earlier real model study of sample, mainly built-in research methods or General/Python, could not be inferred from the fact that individual self-defined experts had improved quality on real investment issues. Nor does the new experience mean that the account is allowed to install the executing specialist kit at will, to provide an independent budget to an expert, or to give the proposed text a trade authorization.
This is without prejudice to the interpretation of an already established engineering judgement: the individual research methodology can be used as a version of the data, replicating the existing Child Implementation and Results System; it is therefore unnecessary to add an Extra Manager with all permission.
Concluding remarks
The simplest step to achieve a custom specialist is to provide a Prompt input box. What makes it truly sustainable is that the method is fixed, the job description can be modified, the actual contribution can be reread, and not “what the user likes” becomes “what the system has done”.
The ETF Agent approach does not require each new individual research methodology to maintain an additional model example, budget, mandate or mission book. Users can build up their research habits over time; the system records in each execution: who uses whom, which version is used, what is actually done.
The experts deserved to be reused and the fact of implementation more so.
Only when the two are separated can a user-defined expert evolve from a one-time role-playing to a capacity that can be upgraded on a continuous basis without bringing down the existing Runtime.
Extension reading and communication
Subagent: Segregation from context to professional collaboration • Models understand, code authorizes ETF Agent Technical Articles repository (welcome Star).
Original language and communication
Read the original ETF Agent Engineering Collection (welcome Star).
Auto-fetched and translated by AI Atlas. All rights belong to the original author.