启动了三个专家，收到了三份报告，最后的答案却仍然像主 Agent 自己写的。此时说“多 Agent 协作成功”，到底成功在哪里？
困难不只是专家有没有运行。我们还需要知道：交回的是什么，主 Agent 实际拿到了什么，答案中的哪些判断与这些发现对应，以及多出来的工作是否值得。
专家被启动、成果被提交、文字进入主 Agent、判断被采用、最终答案变好，是五个不同的问题。 本文讨论怎样逐层取证，而不是再设计一套专家编排框架。
一、先别用“参与次数”回答“贡献多少”
在 2026 年 10 月 30—31 日的一组研究测评中，前 59 次已经结束的运行里，53 次至少有一次带执行身份的 Reviewer 准入，12 次至少有一次普通专家准入。
这两个集合可以重叠，不能相加得到“65 次专家研究”。Reviewer 是复核角色，普通专家承担其他局部工作；准入记录说明某个执行体获准进入流程，不代表它已经完成任务。
继续查最终成果，48 次运行中有至少一份可读最终提交，共 75 份。按同一执行身份关联，其中 62 份来自 Reviewer，13 份来自普通专家。
到这里，我们可以说“哪些专家留下了可读成果”，还不能说“哪些成果被采用”。一个子代理完全可能认真研究并提交，但 Main 没有读到、没有继续读原文，或者有理由不采纳。
因此，统计协作时，运行、执行体、提交和模型请求必须各有自己的分母。否则，几个好看的数字会把同一次工作算成多份贡献。
二、把贡献拆成可核对的交接
这里把主代理称为 Main，把承担局部工作的子代理称为 Child。实际关联应从它们已有的执行身份和结果引用出发，而不是先拿两段文字算相似度。
它不是一张必须逐格通过才能展示正文的审批表。专家没有被采用，可能是合理的研究判断；某项证据无法采集，也不能因此隐藏其他已合法交付的内容。
这张表的用途，是限制我们能作出的结论，并把下一步调查落到正确位置。
三、文字匹配能证明到哪一步
为了检查交接，我们对每份可读最终摘要最多抽取 20 个、每个至少 40 字符的片段，再到该提交之后捕获的 Main 输入中查找。
75 份最终提交里，70 份至少有一个候选片段匹配，涉及 45 次运行。这个观察有价值：至少一些专家文字确实进入了主 Agent 的后续输入，而不是只留在子代理自己的历史中。
但把它叫作“93% 专家采纳率”就错了。
首先，一段文字进入上下文，不代表模型据此改变了判断。其次，命中一个片段不代表整份摘要、证据和限制全部到达。最后，文字没有匹配也不能证明没有交接：结果可能以引用形式进入，或者经过了未被该方法覆盖的改写和选区。
更稳妥的说法是：这是一项有明确采样方法的输入覆盖观察，不是语义采纳或答案质量评分。
还要防止另一个误归因：Main 可能在专家提交前就见过同一来源。即使后来答案和专家措辞一致，也不能据此认定这句话是专家独立贡献的。关联身份与请求先后可以缩小范围，不能自动识别模型内部的因果。
四、用一个费用核查问题说明差别
下面是教学案例，不是上述测评的原始用户资料。
Main 要比较两只基金的持续费用。Child 查到某个费率只适用于特定份额类别，而且不包含一次性交易费用，交回了费率、适用日期、原文引用和这两项限制。
情形一：Main 只拿到结果引用。 这证明材料有一个可继续读取的入口，不能证明本轮已收到费率正文。应继续核对是否发生了获准的读取，以及读取覆盖了什么。
情形二：费率和限制进入了 Main 输入，但答案只保留费率。 传输链路可能成立，答案表达却扩大了适用范围。需要修的是研究综合或表达，而不是再给 Child 增加一个重试。
情形三：答案按份额类别比较，并保留费用范围。 可以建立特定专家发现与最终主张之间的可读对应。但 Main 也可能从相同原始材料独立得出这个判断；要宣称专家带来了质量增益，仍需要可比实验。
三种情形不该被一个 used=true 覆盖。
五、一个不越权的贡献记录应该长什么样
下面是建议用于人工审阅的诊断记录，字段是教学设计，不代表系统已经实现了自动语义贡献判定：
合成诊断记录可以说明：子代理提交可读，主代理输入观察到片段，而主张审阅尚未完成、增量收益尚未测量。
记录应优先引用已有的运行、执行体、提交和请求身份，不另建一套能改写业务事实的“贡献账本”。内容关联可以帮助审阅，但不能签发权限，也不能代替最终答案评估。
先核对子代理提交的执行身份，再检查后续主代理输入：可能观察到正文片段、仅观察到引用，或暂时无法确认。仅有引用时继续核对获准读取，结果可能确认正文或仍未知。观察到正文后独立审阅答案主张，记录对应关系或不一致。
上述三个输入分支分别表示已观察到正文片段、只观察到引用、或证据不足。只有取得实际读取内容后，才把引用分支推进到正文覆盖。最后的主张审阅仍是独立工作；发现不一致要记录，不反向篡改专家原提交。
这种拆分也能减少错误修复。如果 Child 已经交付而 Main 请求没有携带材料，查交接与投影；材料已经到达但结论没有保留限制，查综合和引用；专家做了大量重复劳动，查委派策略与增量价值。
六、怎样比较“多专家是否值得”
到了收益层，不能再用启动数和片段命中率代替对照。
一个可行的后续方案，是固定用户问题、来源版本、研究时点、模型配置与总预算，分别运行单 Main 和多专家方式，对最终交付答案做盲评。记录事实支持、关键遗漏、口径错误，以及端到端时间、全部物理尝试和未知费用。
这里比较的不是哪边专家更多，也不是哪边输出更长。它应该回答：用户真正关心的错误是否减少，重要限制是否保留，多出来的资源是否换来了有价值的信息。
总预算要事先说明。相同每代理预算会让多专家自然获得更多总资源；相同总预算则更接近“同样成本下怎样分工”。两种研究问题都可以成立，但不能把前者结果包装成后者的效率优势。
失败和未完成样本也要保留。只挑顺利完成的多专家案例，再与普通 Main 的全部任务比较，会把运行选择偏差当成架构收益。本文所依据的观察尚未完成这种质量与成本对照。
结语：先证明交接，再讨论贡献
子代理最先提供的是工作过程的隔离，以及局部成果回到主代理的能力。它是否提供了额外判断价值，需要更细的证据。
下一次评审多 Agent 系统，与其问“用了几个专家”，不如选一条关键结论，沿着 原提交 → 后续 Main 输入 → 最终主张 → 原始来源 回查一遍。链路在哪一层断了，就在哪一层改；链路已经成立，也不要急着把它写成收益证明。
输入覆盖是工程事实，语义采用是内容判断，增量价值是比较问题。 把三者分开，才可能判断哪些协作值得保留，哪些只是增加了成本和篇幅。
延伸阅读：Subagent 的上下文隔离与专业化 · 可观测全景中的 Trace 完整性 · 一份报告的成本归属
更多案例见 ETF Agent 技术文集。欢迎为文集点一个 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。
Three experts were activated, three reports were received, and the final answer was still written by Lord Agent himself. This is the word “Mo Agent Success”, where are the successes?
The difficulty is not just that experts have not run. We also need to know what the return is, what the Lord Agent actually got, what the judgement in the answer corresponds to these discoveries, and whether more work is worth it.
Experts are activated, results submitted, text entered, judgement used and final answers improved, and five different questions are raised. The paper discusses how to take evidence from one level to another, rather than designing an expert framework.
First, do not answer “how much” with “the number of participants”.
In a group of studies from 30 to 31 October 2026, 59 previous operations were closed, 53 at least once with executive status, and 12 at least once with regular experts.
These two pools can overlap and cannot be added to “65 expert studies”. The Reviewer is a review role, with ordinary experts performing other local tasks; access records indicate that an implementer is allowed to enter the process and do not mean that it has completed its mission.
The final results continue to be checked, with at least one readable final submission out of 48 operations, totalling 75. Linked by the same implementing status, 62 of them were from Reviewer and 13 from ordinary experts.
Here, we can say “which experts left readable results” and not “ which results were adopted”. A sub-agent may well have been carefully studied and submitted, but Main did not read, did not continue to read the original language or had reason not to adopt it.
Therefore, statistical collaboration must have its own denominator for running, executing, submitting and modelling requests. Otherwise, a few good-looking figures would count the same job as a multi-component contribution.
Disassembly contributions into truncable interfaces
The principal agent is called Main, and the sub-agents who perform part-time work are called Child. The actual connection should be based on their existing implementation identity and reference to results, rather than using two textual equivalents.
It is not an approval form that must be passed on a case-by-case basis to show the text. The expert was not used, which may be a reasonable research judgement; one piece of evidence could not be collected and therefore could not hide other elements that had been lawfully delivered.
The purpose of this table is to limit the conclusions we can make and to put the next step investigation in the right place.
III. What steps does text match to prove?
In order to check the handover, we extract up to 20 segments of each readable final summary, each at least 40 characters, and then search for the Main input captured after that submission.
Of the 75 final submissions, 70 were matched by at least one candidate clip, involving 45 operations. This observation is valuable: at least some of the expert texts did enter the main Agent's subsequent input, rather than simply remaining in the history of sub-agents.
But it's wrong to call it "93% expert adoption."
First, a text enters context, which does not mean that the model changes the judgement accordingly. Second, a single piece of the hit did not represent the whole of summary, evidence and restrictions on full arrival. Finally, the absence of a match in the text does not prove that there is no connection: the result may be in the form of a reference, or the result may have been rewrited and the constituency not covered by the method.
More conservatively, it is an input-covered observation with clear sampling methods, not semantic adoption or answer quality rating.
Another cause of error is also to be prevented: Main may have seen the same source before the experts submitted their reports. Even if the answer were later to be consistent with the expert's wording, it would not be possible to conclude that the statement was an independent contribution by the expert. Related identities and requests can be narrowed down and the causes and consequences within the model cannot be automatically identified.
IV. A cost verification issue to illustrate differences
The following are teaching cases and are not original users of these assessments.
Main compares the continuing costs of the two funds. Child found that a rate applied only to a specific share category and did not include one-time transaction costs, and returned the rate, the applicable date, the original reference and both restrictions.
Case one: Main only gets the results quoted. This supporting material has a continuing readable entry point and does not support the receipt of the original rate for the current round. Checks should continue to be made on whether approved readings had taken place and what was covered.
Case II: Rates and restrictions entered Main but only rates are retained. The transmission chain may be established, while the expression of the answer extends the scope of application. What is needed is research synthesis or expression, rather than adding a retry to Child.
Scenario III: The answer is compared by share category and the cost range is maintained. Readable correspondence could be established between a particular expert ' s findings and the final claim. Main may, however, independently draw this judgement from the same source material; comparable experiments are still required to claim that the experts have brought qualitative gains.
Three situations should not be covered by a used=true.
V. What a record of non-excessive contributions should look like
The following are recommended diagnostic records for manual review. field is a pedagogic design and does not represent an automatic semantic contribution determination of the system:
Synthetic diagnostic records show that sub-agents submit readable, the main agent enters observed footage, while claims for review have not been completed and incremental gains have not been measured.
The records should refer, as a matter of priority, to the operational, executing, submitting and requesting identities available, and should not create a separate “contribution book” that can be adapted to business facts. Content connections can help with review, but cannot issue permission or substitute for a final answer assessment.
Check the executing identity submitted by the sub-agent and then check the subsequent master agent input: it may be possible to observe the body section, only the reference, or it may not be confirmed for the time being. Only when quoted are checks continued for permitted reading, the results may confirm the text or remain unknown. An independent review of the answer claims after the body has been observed to record correspondence or inconsistencies.
Each of the three input branches indicated that the body parts had been observed, only cited or insufficient evidence. Only when you actually read the content will you advance the reference branch to the body of the text. The final review of the claim remains an independent exercise; inconsistencies are found to be recorded and not to be presented in reverse to tampering with the original expert.
This split can also reduce error repair. If Child has been delivered and Main ' s request is not accompanied by materials, handing over and projecting; the material has arrived but the conclusion is open-ended, consolidated and quoted; and experts have done extensive duplication work to verify assignment strategies and incremental values.
How to compare “the value of multiple experts”
At the level of the proceeds, no comparison can be replaced by a start-up number and a fractional hit rate.
A viable follow-up option is fixed user questions, source versions, time points for research, model configuration and total budget, operating in single Main and multi-expert formats, respectively, with blind feedback on the final delivery answers. Recording of factual support, critical omissions, calibration errors, end-to-end time, all physical attempts and unknown costs.
It's not more expert than expert, it's not more output. It should answer whether there had been a reduction in the number of errors of real concern to users, whether important restrictions had been retained and whether the additional resources had been replaced by valuable information.
The total budget must be stated in advance. The same agent budget would naturally increase the total resources of many experts; the same general budget would be closer to “how to divide at the same cost”. Both research issues can be established, but the results of the former cannot be packaged into the latter ' s efficiency advantages.
Failure and failure to complete sample are also to be retained. Only a successful multi-expert case and a comparison with normal Main's full tasks would see the operational selection bias as a structural gain. The observations on which this paper is based have not yet completed this quality versus cost.
Concluding remarks: Prove handover and discuss contributions
Sub-agents first provide segregation of work processes and the ability to return local results to the principal agent. Whether it provides additional value for judgement requires more detailed evidence.
The next review of the Agent system, instead of asking "a few experts were used", choose a key conclusion, following the original submission of a follow-up Main input that ultimately claims that the original source is back. The chain is changed at which level it is broken; the link has been established and does not need to be written as a certificate of profit.
The input cover is a work fact, semantics are judged by content and incremental value is a comparative issue. The separation of the three would make it possible to judge which collaboration was worth retaining and which merely increased costs and space.
Extension: The Subagent's context Segregation and Specialization Trace Completeness in a Visible Panorama Costing of a Report
More examples can be found in the ETF Agent Technical Papers. Welcome to a collection for Star.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.