用 AI 智能体处理长周期金融工作

我们正在 2026 年纽约 AI 工程师大会现场，我身边是来自 Arca 的 Aman Singh。Arca 最近为一家 AI 原生财富管理公司融资 6400 万美元。那么首先，Aman，你能跟我们讲讲你在 Arca 的角色、你们提供的金融服务，以及你们如何使用 AI 智能体来运营这些服务吗？

We're here at the AI Engineer New York conference 2026 and I'm here with Aman Singh from Arca. Arca's recently raised $64 million for an AI native wealth management firm. So first of all, Aman, can you tell us a bit about your role at Arca and the financial services you provide and how you're using AI agents to run them?

当然可以。感谢邀请。我在 Arca 的角色是领导 AI 平台团队。我们正在构建底层的 harness 以及用 AI 运营业务的能力，这是我们的终极目标。就我们为客户提供的服务而言，这是一系列广泛的金融服务。包括财务建议、资产财务规划，以及随着人们变得更富有而涉及的税务规划、保险和遗产规划。比如你可能刚卖掉一家公司，涉及各种合作伙伴等很多复杂情况，对吧？我们可以帮你理清这些。我们也有客户是大型科技高管，他们试图制定真正的提前退休计划，对吧？还有小家庭，比如想买第一套房子。

Yeah, for sure. Thanks for having me. My role at Arca is I lead the AI platform team. So we're building the underlying harness and our capabilities to run the business with AI, which is our ultimate goal. In terms of the services that we provide to our clients, this is a broad spectrum of financial services. So that includes financial advice, financial planning for their assets, and tax planning and insurance and estate planning as people get wealthier. Maybe you are someone who just sold a business for example and there's lots of complexities with all the different partners etc, right? So we can help you think through that. We also have clients that are really large technology executives that are trying to figure out a real plan for retiring early, right? And also small families that are trying to buy their first home for example.

那么你们目前的客户，你需要跟他们解释使用智能体、为什么使用智能体，并告诉他们一些背后的技术吗？

So the clients that you have so far, do you have to kind of talk them through using agents and why you're using agents and tell them a little bit about the technology behind it?

是的，AI 在 Arca 的角色不是取代顾问，而是取代顾问周围的所有工作。所以我们的客户并不直接与 AI 交互。他们体验到的是一种行业内前所未有的服务水平，通过人类顾问的共情和关怀来传递。我们正在为每位客户构建背后的机制，它不断思考市场在做什么，以及它如何影响 Richard？你生活中还有什么其他事情？你生活中哪些方面会影响你的财务？我们能够更加主动、更加响应迅速，并通过 AI 完成所有后台工作，提供更好的客户服务水平。

Yeah, so the role of AI at Arca is not to replace the advisor. It is to replace all of the work around the advisor. So our clients don't really interface with AI directly. All they experience is a level of service that has not been possible in this industry before, delivered through the empathy and care of a human advisor. So we're building the machinery behind every client that's constantly thinking about what is the market doing and how does it impact Richard? What else is going on in your life? What are the textures of your life that impact your finances? And we're able to be way more proactive, way more responsive, and just deliver a better level of client service using AI that does all of the back office stuff.

好的。那我们谈谈一些技术方面。你们使用哪些模型来驱动智能体，以及如何决定哪个模型最适合哪个操作？

Okay. So let's get into some of the technical aspects of it. Which models are you using to power your agents and how do you decide which model is best for which action?

是的。今天我们使用的是来自前沿实验室的前沿模型，比如 Anthropic 和 Open AI。Claude 一直是我们的主要模型，但多年来我们经历了多次模型迭代。我们曾经用 Gemini，然后是 Open AI，再然后是 Anthropic。指导我们所有决策的是我们内部的基准测试，针对我们通用 harness 在运营业务中必须展现的各种能力。比如最近，尽管外部基准测试声称 Luna 很棒，但对我们来说并没有像预期那样奏效。它太慢了，而且成本很高。所以我们的内部基准测试指导我们的模型决策，原则上我们希望保持模型无关，不依赖某个提供商或这些实验室的某个 harness，并为我们保持这种可选性。

Yeah. So today we're using the frontier models from the frontier labs like Anthropic and Open AI. Claude is kind of our primary model has been but we've gone through several iterations of models over the years. We used to be in Gemini way back when and then Open AI and then Anthropic. And what's guiding all of our decisions are our benchmarks that we have internally for all of the various capabilities of our general harness that it has to express in running our business. And so just recently for example, despite the external benchmarks claiming great things for Luna, it didn't really quite work out for us as well as we had expected. It was way too slow and it was costing a lot. And so our internal benchmarks kind of guide our model decisions and as a principle we want to be model agnostic and sort of not take a dependency on a certain provider or certain harness from one of these labs and kind of keep that optionality open for us.

好的。你刚才提到了 harness。那么你们是否根据使用的模型不同而使用不同的 harness？

Okay. So you just mentioned harnesses. So are you using different harnesses depending on which model you're using?

不，我们内部构建了自己的 harness，它是模型无关的。

No, we have our own kind of built internally that is model agnostic.

好的。那么当财务顾问做税务报告之类的事情时，harness 如何决定使用哪个模型？这是自动化的吗？

Okay. So when a financial advisor is doing something like a taxation report or something like that, how does the harness decide which model to use? Is that automated?

嗯，是的。我们根据任务复杂度的不同，设有不同层级的模型。像协助制定税务计划或进行税务分析这样的任务，会使用我们最重层级的模型。在该层级内，有一定的灵活性，所以你可以使用 Opus 层级的模型，或 Sonic 层级的模型，或重型 GPT 模型。但基于我们的智能体参与的大量后台活动，对吧？这些活动根据任务复杂度使用较低层级的模型。这就是我们决定哪个模型服务哪个任务的方式。

Well, yes. So we have certain tiers of models based on different complexities of tasks. So a task like assisting with a tax plan or doing a tax analysis would get our heaviest tier of models. Inside of that tier, it is a little bit flexible, so you could use an Opus tier model or a Sonic tier model or a heavy GPT model. But based on there's a lot of background activity that our agents are participating in, right? That is using a lower tier of model based on the complexity of the task. So that's kind of how we make decisions on which model serves which task.

好的。那么是财务顾问在使用产品时实际做出那个决定吗？

Okay. So is it the financial advisor that's actually making that decision while they're using the product?

不完全是。他们并不知情。我们已将此完全抽象化，不让他们接触。他们只是像平常一样与我们的产品交互，而那些决定由我们内部做出。

Not really. They're not aware of the happening. We've abstracted this away from them entirely. They kind of just interact with our product as they would, and those decisions are made internally by us.

好的。那么在你今天早些时候在会议上的演讲中，你谈到了一个税单的例子。我想知道你是否可以带我们走一遍调整客户投资组合的例子，从识别需求到可能重新平衡投资组合再到执行交易。你能跟我们讲讲这个过程中哪些部分由智能体完成，哪些部分由人类顾问完成吗？

Okay. So in your presentation at the conference earlier today, you talked about a tax bill example. I just wondered if you could maybe walk us through an example of adjusting a client's portfolio, investment portfolio, from identifying a need to maybe rebalance the portfolio to executing trades. Can you talk us through what parts of that process the agent does and what parts the human advisor does?

对。就像演讲中那样，这件事会跨越 10 天。所以智能体最终负责交付那个结果，但有些部分它们做，有些部分它们不做。在这个例子中，当客户带着整个税单发邮件给我们时，有一部分分析是：这位客户的钱今天分配在哪里？以及以最优方式清算部分资金以产生 10 万美元？这最适合智能体作为分析任务来承担。

Right. So just like in the talk, this is a thing that will span 10 days. And so agents are ultimately responsible for delivering that outcome, but there are certain parts that they do and certain parts that they do not. So in this example, when the client emails us with this whole tax bill, there's a piece of the analysis of like, where is this client's money allocated today? And what is the most optimal way to liquidate some of that to produce $100,000? That is best suited for agents to take on as an analysis task.

对。

Right.

所以它们就做这个。

So that they do exactly that.

不过它们是否有某些限制，比如它们不能以正确金额执行交易？

Do they have certain limits though that they can't, you know, they can't execute a trade at correct amount?

所以分析部分是全自动的，但它们无法自动执行任何资金移动或清算。我们与托管人有非常深入的集成，实际上提供了 API 来做到这一点，但我们有严格的护栏，对于涉及移动资金或清算资产，甚至直接与最终客户沟通的事情，都需要人类在环，因为我们期望一定的服务和质量水平。所以所有内部的分析将由智能体完成，但它们也会在需要时拉入合适的人来执行交易。

So the analysis part is fully automated, but they're unable to actually execute any sort of money movement or liquidation automatically. We have a very deep integration with our custodian that actually gives us APIs to be able to do that, but we have strict guardrails around human in the loop for things that would either move money or liquidate assets or even directly communicate with our end clients because there's a level of service and quality that we expect. So all of the internal analysis will be done by agents, but they will also pull in the right person to be able to execute the trade when needed.

你提到这些流程往往要 10 天，或者至少比现在大多数智能体的运行时间更长。那么当一个智能体在开始任务几天后才恢复执行时，你如何恢复——如何确认这个计划仍然符合客户的需求？

You mentioned that these processes tend to be 10 days or at least a longer running time than most agent stuff these days. So when an agent resumes a task maybe days later than they started it, how do you restore — how do you check that the plan still fits the client's needs?

好，这个问题很有意思。我们有大量这类信号，就像我在演讲中描述的那样。比如你可以想象有一个邮件线程正在收到回复，有人在里面讨论，对吧？任何新增的回复都会交给同一个正在阅读该邮件线程并做出反应的智能体。但一个全新的信号——比如中途打进来的一通电话——是一个全新的信号，不依附于任何正在进行的流程。于是会有一个新的智能体被唤醒，去弄清楚这意味着什么。而它们理解当下正在发生什么的方式，是我们的内部记忆系统。所以我们为每个客户维护这种持久状态：现在正在发生什么、我们要服务什么、谁负责什么、到目前为止发生了什么、我们在追踪什么后续事项——接下来几天需要立即做什么、长期在追踪什么、哪个人负责哪个动作。所有这些都由我们的记忆系统追踪，我们的智能体也会向其中贡献信息。但同时还有一个类似“做梦”的过程，它会反思与客户 Richard 相关的世界状态，并能够更新这些记忆系统，以持久化长周期任务。所以那个收到新信号的智能体会迅速了解客户当前的情况，它会知道：“好，我们实际上正在执行一笔大额税单。”它也能看出：“好，我的信号和那笔税单有关，那我来为那个世界的状态做点贡献。”

Yeah. So this is an interesting one to think about. We have tons of these signals, like I described in the talk. Some are, for example — you can imagine there's an email thread that is receiving responses and there's chatter happening, right? Any additional response to that goes to the same agent that's reading that email thread and reacting to it. But a fresh signal — like a phone call that was made in the middle — is a brand new signal that isn't attached to any other ongoing process. And so a fresh agent wakes up to figure out what this means. And the way they understand what is happening right now is our internal memory systems. So we have this persistence for every client: what is happening right now, what are we trying to service, who owns what, what has happened so far, what are we tracking for next — what needs to happen immediately in the next couple of days, what are we tracking for the long term, which individual is responsible for which action. All of this is being tracked by our memory systems that our agents contribute to. But also, there is a kind of really dreaming process that reflects on the state of the world as it relates to Richard, the client, and is able to update these memory systems to persist long-horizon tasks. So that agent that received a new signal quickly ramps up on what the situation is with the client, and it'll know, "Okay, we're actually executing a big tax bill." And it'll be able to see, "Okay, my signal relates to that tax bill, so let me contribute to the state of that world."

你能不能谈谈你在产品里构建或接入的一些安全方面的机制，显然是为了确保过程中不会损失任何资金？

Can you talk a little bit about the safety aspects that you've maybe built onto your product or harnessed, to make sure obviously that no money is lost along the way?

好。我们所有的安全限制都是完全确定性的，它们把决策权从智能体手中拿走。比如某些工具允许智能体自己执行。如果是简单的记录类工具，它就会执行。另外，每当调用会写入数据的工具时，我们都会对该动作运行基于 LLM 的验证。这些是快速的 LLM 调用，会分析对我们数据库的每一次写入。所以这是我们运行的又一层验证。然后我们还会摄入大量外部信息。比如进来的邮件、进来的文档、打进来的电话，这些都是提示注入的目标。所以我们在边缘运行验证，检测某些类型的提示注入是否存在。然后任何时候出现数据离开我们系统的情况——比如最明显的网页搜索，或者我们给 Richard 发日历邀请，对吧？那是外部数据——我们既有确定性检查，检测我们已知属于 PII 的某些信息是否存在，也有小模型的共识机制，确保没有任何数据外泄的企图。所以我们在 harness 里内置了大量不同的护栏，原则是我们希望在能力的边缘进行遏制，而不是更早。

Yeah. So all of our safety restrictions are fully deterministic and they take the decision-making off the agent's hands. So for example, certain tools let the agent execute the tool itself. So if it's a simple record-keeping kind of tool, it will execute it. We also — every time a tool is called that writes data, we have LLM-based verification running on that action. These are fast LLM calls that analyze every single write to our database. So that's another layer of verification we're running. And then we also ingest a lot of external information. So there are emails coming in, documents coming in, phone calls being made, which are targets for prompt injection. So we run verifications at the edges that detect the presence of certain flavors of prompt injection. And then anytime there is a case where data leaves our system — for example, in a web search in the most obvious case, but also if we're sending a calendar invite to Richard, right? That's external data — we have both deterministic checks on the presence of certain information that we know is PII, but also a consensus of small models that ensure that there isn't any kind of data exfiltration being attempted. So we have tons of different guardrails built into the harness, with the principle that we want to contain them at the edges of capability and not earlier than that.

而且因为你们身处金融行业，在使用技术方面是否有某些必须遵守的监管要求？还是说更多是关于结果以及你们与客户打交道的方式？

And because you're in the financial industry, are there certain regulations that are required that you have to keep to in terms of the technology that you're using? Or is it more about outcomes and the way you deal with clients?

对，行业关心的是结果，但合规关心的是可审计性，对吧？所以我们的系统做到默认可审计。智能体采取的每一个动作都会产生变更日志：改了什么、谁改的、来自系统的哪个部分、什么时候改的。顺便说一句，这不仅适用于智能体，也适用于人类操作者。所以如果你在 UI 里改了影响客户的东西，我们也会记录该动作。这让我们既能默认可审计，也支撑了我们的一些智能体能力——我们追踪的 UI 变更同样可以用来告知智能体发生了这件事。所以可审计性一直是我们叙事的重要部分，既是为了符合监管要求，也是为了我们的智能体循环。

Yeah, the industry cares about outcomes, but compliance cares about auditability, right? And so our system is such that auditability is by default. Every single action that an agent takes produces a change log of what changed, who changed it, from which part of the system, and when. And this applies not just to agents, by the way. It applies to human actors as well. So if you change something in the UI that affects the client, we'll log that action. So it lets us both get auditability by default, but also powers some of our agentic capabilities, where the same thing that we track like a UI change also could be used to inform the agent that that happened. So auditability has been a big part of our story, both for compliance with regulation, but also for our agentic loop.

你在演讲里有一句话很妙：“智能体提示人的次数，比人提示智能体的次数更多。”那么智能体如何决定让谁参与进来，以及它们需要向那个人提供多少上下文，才能让对方完成需要做的事？

In your presentation you had a nice line: "Agents prompt people more than people prompt agents." So how do the agents decide who to involve, and how much of the context do they need to provide that person in order to do what they need to do?

好，这又是一个有意思的问题。这和我们行业里人的工作方式类似。我们追踪一个为期 10 天的流程，涉及多个参与者。第七步的参与者可能还不需要知道这件事，对吧？他们会在合适的时间被拉进来。所以我们的智能体凭借长期记忆持久化，加上每天对正在发生之事的反思，会反思当下世界状态处于何处。当出现某个信号表明必须把某个人拉进来时，它们能够在那个时刻拉进正确的人。而这一切都是我们做的工作，把行业运作方式的上下文操作化。但不仅如此，还有 Arca 希望它如何运作，以及我们对这个行业工作如何发生的原则。这也是我们不只是一家 SaaS 公司的另一个优势。我们其实不是 SaaS 公司，我们自己就是那家事务所。所以我们可以把这项工作本身的发生方式标准化，让我们的智能体能够以工作必须发生的标准方式参与其中。

Yeah, another interesting one. So this is similar to how humans work in our industry. There is a process that we're tracking over 10 days. There are multiple actors involved. The actor at step seven may not need to know about this just yet, right? They're pulled in at the right time. And so our agents, with our long memory persistence plus the daily reflection on what is happening, reflect on where the state of the world is at the moment. And they're able to pull in the right person at that time, when there's a signal that implies that person must be pulled in. And all of this is work that we have done to operationalize the context of how this industry works. But not just that — how does Arca want it to work? And what are our principles of how work happens in this industry? And that is another advantage of us being not just a SaaS company. We're actually not a SaaS company. We are the firm ourselves. So we can standardize the way this work even happens, so that our agents can participate in standard ways that work must happen.

那么在会议早些时候，Anthropic 办了一场关于他们托管智能体基础设施的工作坊。其中很多部分——甚至编排——现在都变得更自动化了。

So earlier in the conference, Anthropic did a workshop about their managed agents infrastructure. And a lot of it — even the orchestration — is becoming more automated now.

那么，你预见到会使用那种功能吗？因为它显然带来一种危险，即人类不再总是处于循环之中。所以，你有没有想过用那种技术来自动化更多流程？

So, do you foresee kind of using that type of functionality cuz it obviously it presents a danger in that the human is not necessarily always in the loop anymore. So, have you had any thoughts about using that technology to automate even more of the process?

嗯，我认为我们非常看好 Anthropic 以及像 OpenAI 这样的实验室的模型。嗯，但对于依赖像托管智能体这样的产品，我们没那么看好。嗯，因为托管智能体的一个关键点在于模型如何持久化并贡献关于世界正在发生什么的记忆。嗯，以及它们成功所需的所有操作知识。嗯，所以，让他们拥有那部分业务是一种风险，因为拥有那部分就是拥有你的智能，这能让你有能力影响业务。如果你把智能交给另一个提供商，你就会失去一些控制权，而你需要这种控制权来构建独特且高度定制化的东西，从而真正影响你的业务。

Um, I think we are very bullish on the models from Anthropic and labs like OpenAI. Um, but less so on taking a dependency on a product like managed agents. Um, because managed agents, one of their key things is how the models kind of persist and contribute memory as to what is happening in the world. Um, and all of the operational knowledge that they need to use to be successful. Um, and so, letting them own that part of the business is kind of like a risk because owning that is owning your intelligence is kind of gives you the ability to impact the business. If you hand off the intelligence to another provider, you lose some of that control that you need to build something that is unique and tailored to your business to such a great degree that it can actually impact your business.

就 Anthropic 而言，他们最近也发布了金融服务。所以，他们正在与一大堆不同的公司合作。你认为他们所做的与你们试图做的有竞争性吗？还是你认为它们是不同的东西？

And just on Anthropic, they've also recently released financial services. So, they're partnering with a whole bunch of different firms. Do you think what they're doing is kind of competitive to what you're trying to do or do you see them as different things?

是的，我认为 Anthropic 进入我们这个行业是对我们行业的极大肯定。但我思考的方式是，一家拥有 AI 工具的服务公司与一家天生就是 AI 原生的公司是非常不同的。拥有 AI 工具的公司需要用户提示智能体来产出工作或完成工作。但在这里，就像我在 Arca 说的，智能体提示人多于人提示智能体，它是一种主动式服务，为他们 24/7 运行，并且以我们期望的最标准方式运营业务，达到我们渴望的服务水平，如果你是一家拥有 AI 工具的公司，那看起来就非常不同。

Yeah, I think Anthropic getting into our kind of industry is very validating for our industry. But the way I think about it is that it is very different to be a services firm that has access to AI tools versus a firm that was built to be AI native. The firm with AI tools requires their users to prompt the agents kind of to produce work or to do work. But here like I said at Arca agents prompt people more than people prompt agents and it is a proactive kind of service that is running for them 24/7 and it's kind of like running the business in the most standard way that we expect at the level of service that we aspire to which looks very different if you're a company with AI tools.

是的。是的。好的。那么再多谈一点安全方面。你能解释一下你们流程中哪些部分在沙盒中运行，以及你们如何在任何人行动之前验证对财务计划的更改吗？

Yes. Yes. Okay. So a little bit more about on the security side. So can you explain like what in your processes runs inside a sandbox and how do you verify like a change to a financial plan you know before anyone acts on it?

是的。我们使用沙盒的方式是，沙盒是我们框架的一个组件。我们的智能体实际上并不在沙盒内部。它使用沙盒来卸载计算，我们已经构建了系统，使沙盒内的代码可以安全地访问我们的数据。所以我们有这样一种机制，将它编写的代码转换为我们所期望的严格格式，丢弃代码，并在其到达后端之前对这种格式施加护栏。所以我们允许智能体使用代码查询我们的数据，这意味着它们可以聚合数据、处理数据，并按照它们认为合适的方式使用数据来生成像财务计划这样的东西。在同一个沙盒内，它们生成这些漂亮的工件，我们可以与客户分享，比如一个 PDF，上面写着这是你的财务计划，这是我们计划如何从财务上处理你的生活。一旦生成，我们就运行验证器。所以，从沙盒中产生的是某种类型的工件。一旦生成，我们就会对其进行一系列不同的验证。一些是确定性的检查，比如“嘿，你是不是刚把一个原始账号放在那里了？你不应该那样做。”所以，我们可以确定性地检查这一点。但也会用快速模型，比如较小的模型或 Jev 来分析是否存在可能使此工件不合规的内容。例如，你真的不应该在该工件中承诺回报。所以，这是我们在它可发布之前会检查的另一件事。关于这一点我要说的最后一点是，这样做实际上使我们的合规性比人类做的更好，因为 AI 不知疲倦地发现其中哪怕是最微小的不合规文字。所以，是的，我们将继续在这方面投入更多，并在能力边缘进行约束。

Yeah. So the way we use sandboxes is that the sandbox is one component of our harness. Our agent isn't actually inside of the sandbox. It uses the sandbox to offload computation and we've built systems to make it safe for code inside of that sandbox to access our data. So we kind of have this thing which translates the code it wrote into a strict format that we expect, throw away the code and you kind of instill guardrails on that format before it reaches your back end. So we allow the agents to use code to query our data which means that they can aggregate it and they can munch it they can kind of work with it as they see fit to produce such a thing like a financial plan. Inside of the same sandbox they produce this, you know, beautiful artifacts that we can share with clients that has like it's a PDF of like here's your financial plan and here's how we're planning to approach your life financially. And once that is produced is where we run our verifiers. So, it's a certain type of artifact that is produced from the sandbox. The moment that is produced we run bunch of different verifications on it. Some deterministic that check for presence of like, "Hey, did you just put like a raw account number in there? You should not do that." So, deterministically we can check that. But also fast models like the smaller models or Jev to analyze for presence of things that might make this artifact non-compliant. Like for example, you really should not be promising returns in that artifact. So, that's another thing that we will check before it is kind of a publishable. And the last thing I'll say on that is that doing this actually makes our compliance better than if humans were doing it because AI is tireless at finding even the smallest mention of a non-compliant piece of writing in there. So, yeah, we'll continue to invest in more of that and kind of constrain at the capability edges.

是的。你提到了 Jev。所以，你们已经开始在流程中使用决策模型。那么，除了 Jev，你们还在关注哪些其他产品？

Yes. Yeah, you mentioned Jev there. So, you have started to use decision models as part of your process. So, what other products other than Jev are you sort of looking at?

是的，我们注意到当模型处理财务数据时，对吧？当它们回答需要财务知识的问题时，它们对自己拥有的知识非常自信。一个小例子是，模型会很高兴地使用 2025 年的税率等级来进行分析，并说，好吧，就是这样。但这显然是不正确的。所以，这是我们在框架中使用决策模型的方式之一，即检查当模型即将做出财务建议或任何需要财务严谨性的行动时，会检测并推动模型使用我们关于当今世界真实事实的已验证知识库。或者使用它们的网络搜索。所以，这有点像是关于框架和拥有框架的一点，让你能够做这样的事情，帮助你的模型在关键之处保持准确。

Yeah, so what we noticed is when models are working with financial data, right? When they're answering questions that require financial knowledge, they are very confident in the knowledge that they have. So, a small example is that models will very happily use the tax brackets from 2025 to come up with an analysis and be like, all right, that is it. But that is obviously incorrect. And so, that is one of the ways in which we're using decision models in our harness is to check when the model is about to make a financial recommendation or anything any action that requires financial rigor will detect and will nudge the model to kind of use our knowledge base of verified facts about the world that is true today. And or use their kind of web search. So, kind of makes it this is kind of a point about the harness and owning the harness that lets you do things like this that help your models be accurate where it matters.

那么，你们是否在例如交易执行失败后的恢复中使用决策模型？那里有决策模型参与吗？还是有其他路径让智能体重回正轨？

And do you use decision models in for example, recovery if like a trade fails to execute? Is there a decision model involved there or is there another path to get the agent back on track?

不，今天没有。那不是我们今天正在探索的用例。

No, not today. That is not a use case that we're exploring today.

好的。嗯，那么，你知道，如果一步交易确实失败了，例如，在你们的流程中，智能体目前如何恢复？

Okay. Um, so, um, you know, if a step a trade does fail for example, how does the agent recover at the moment in your process?

所以，嗯，目前，我们并没有用智能体进行自主交易。很大程度上是，嗯，智能体会提议我们进行交易。

So, um, so, at the moment, we are not doing autonomous trading with agents. It is very much a um, the agents will propose that we make trades.

我明白了。

I see.

嗯，这完全由我们的专家来决定这是否是正确的交易并执行交易。

Um, and it is fully up to our experts to figure out if that is the right trade to be making and to execute the trade.

好的。那么，沿着同样的思路，嗯，你的幻灯片上说，每一次员工编辑或取消都带有一个 Y，并成为评估。那么，这种反馈如何作为可衡量的改进被转回系统？

Okay. So, uh, on the same sort of line of that, um, your slides say that every staff edit or cancellation comes with a Y and come it becomes an eval. So, how does that feedback get turned back into the system as a measurable improvement?

是的。

Yeah.

这是我们作为业务公司而非 SaaS 公司的另一个优势：我们有权要求他们给我们反馈。我们将反馈循环回改进的方式是你会预期的那样——所有反馈都被导入一个数据存储，该存储将反馈与产生它的智能体的轨迹耦合起来。所以我们能够拿到那个轨迹，把它放在反馈旁边，每天晚上我们运行一个流程，从现实世界中收集我们智能体在野外运行的轨迹，并在上面运行一个元智能体，提出对我们框架的修改建议。所以也许是误导性的工具，或者出错的东西，或者我们的知识库中缺少知识，它们会把这些信息呈现给我们，然后我们就能改进它。今天，在我们实际发布提示更改之前，我们仍然有一个人工工程师在循环中。

That is another advantage of us being the business and not a SaaS company: we have the ability to require them to give us feedback. The way we loop the feedback back into improvements is something you would expect—all of that feedback gets piped into a data store that couples that feedback with the trace of the agent that produced it. So we're able to take that trace, put it next to the feedback, and every night we run this process where we collect traces from the real world, from our agents in the wild, and run a meta agent on top that suggests changes to our harness. So maybe tools that were misleading, or things that were errored, or there's missing knowledge in our knowledge base, and they'll surface this information to us, and then we'll be able to improve it. Today we still have a human engineer in the loop before we actually ship a prompt change.

是的。

Yes.

我注意到，如果你让智能体自己来做，它们会做出非常糟糕的提示更改。每一个小修复都会变成整个提示。它们无法保持元视角，将修复泛化成一个通用的提示。所以我们有来自产品的反馈循环，我们直接作为一等产品功能来收集评估并将其转化为改进。

I've noticed that agents will make really nasty prompt changes if you leave them to do it. Every little fix becomes the whole prompt. And they're unable to stay meta and generalize the fix into a prompt that generalizes. So we have feedback loops from our product that we collect directly as first-class product features to collect evals and turn them into improvements.

嗯。

Hm.

好的。你在演讲中还提到将公司运作的知识转化为智能体上下文，这实际上是许多前置部署工程师尝试做的事情——将组织的内部知识转化为智能体可操作的东西。对你们公司来说，这在技术上是什么样子的，你们如何保持更新？

Okay. You also mentioned in your presentation about turning knowledge of how a firm works into agent context, which is actually something a lot of forward deployed engineers try to do—turn an organization's internal knowledge into actionable things for agents. What does that look like for your company technically, and how do you keep it up to date?

是的。所以我们有一个文件形式的知识库，我们保持更新的方式是把它交给我们的专家来策划和更新。所以它是我们产品中的一等功能。在左侧有一个名为“知识”的标签页，其中包含我们的智能体正在使用的完整操作上下文。当它们处理事情时，它们能够引用知识。所以这就形成了一个循环：哦，我们注意到知识不正确。然后这在我们的轨迹分析中呈现给我们。而且因为它是产品本身的一部分，人们能够直接进去编辑技能。这就是它的运作方式。

Yeah. So we have a knowledge base of files that is represented as files, and the way we keep that up to date is we've handed it to our experts to curate and update. So it is a first-class feature in our product. On the left-hand side there's a tab called knowledge, which has the full operational context that our agents are using. And as they work on things, they're able to cite knowledge. And so it creates this loop of like, oh, we notice that the knowledge is incorrect. And that gets surfaced to us in our trace analysis. And because it is a part of the product itself, people are able to go in and edit the skill directly. That's how it kind of works.

好的。你提到你们实际上通过这些服务向客户提供——我假设你们公司有财务顾问团队。那么，你能给我们举一个例子吗？一个通过智能体系统在投资或税务等方面有可衡量改善的客户，也许还有一些关于智能体架构中哪些部分帮助该客户获得如此好回报的细节？

All right. Okay. So, you mentioned you're actually offering these services to clients through—I'm assuming you have a staff of financial advisers in your firm. So, can you maybe give us an example of a client that you've had that's had a measurable improvement in their investments or taxation or whatever through the agent system, and maybe some details about what in the agent architecture helped that client get such a good return?

是的，我给你一个最近的例子。这个人还不是客户，但正处于成为客户的阶段。所以我们有了他们所有的文件、对账单、当前投资等。这个人已经在与一位顾问合作。所以他们与我们见面，告诉我们：嘿，我已经有这个顾问了。他们收取我 1,500 美元的服务费，而且对我不是很响应。所以我在和你们聊。我们的智能体已经读完了所有这些文件，分析了它们，分析了这个人陈述的目标，对吧？他们还分析了这位当前顾问是谁的对账单，以及他们收取了多少费用。他们发现这个人收取了一些他们不知道的费用。并向他们揭示了这一点，让他们不仅意识到自己被多收费了，还意识到那个人甚至不是受托人。所以他们没有义务为他们的利益行事。这就是我们的智能体在分析财务状况时带来的那种对细节的关注，如果没有这些，就不可能做到。有几个主动性的例子，我们的智能体产生的，比如某人有一个截止日期快到了，或者他们有错过 RMD 的风险，或者类似的事情。所以，这在规模上对人类来说很难跟踪，但这是我们 Arca 智能体的默认行为。

Yeah, I'll give you a recent example. This person wasn't a client yet, but they were at the point of becoming a client. So we had all of their documents, their statements, their current investments, etc. And this person was already working with an adviser. So they meet with us, they tell us that, hey, I have this adviser already. They're charging me $1,500 for their services, and they're not very responsive to me. So I'm chatting with you. And so, our agents have read through all of those documents, they've analyzed them, they've analyzed what the person's goals are that they state, right? They've also analyzed the statement of like who this current advisor is, and how much they're charging them. And they discover that there are fees that this person was charging them that they weren't aware of. And surfaces this to them, and makes them not only aware of the fact that they're being overcharged, but also aware that that person wasn't even fiduciary. So they're not obligated to act in their interests. So this is the sort of attention to detail that our agents bring about when they're analyzing a financial picture, that sort of would not have been possible if we didn't have these. There's several examples of proactivity that our agents produce, like someone has a deadline coming up, or they're at risk of missing their RMD, or that kind of thing. And so, this is difficult at scale for humans to keep track of, but it is the default behavior of our agents at Arca.

是的。太棒了。嗯，这可能是停下来得好时机。所以，非常感谢你，Aman，参加我们的节目。

Yeah. Wonderful. Well, that's probably a good point to stop at. So, thank you very much, Aman, for joining us.

感谢邀请我。

Thank you for having me.