Latent Space
Agents Prompt People: Using Agents for Long-Horizon Financial Work
00:0000:26:05
00:00
We're actually not a SaaS company, we are the firm ourselves.
我们其实不是一家SaaS公司，我们自己就是那家公司。
00:03
Agents prompt people more than people prompt agents
Agent主动提示人的次数比人提示Agent的次数还多
00:06
and it is a proactive kind of service that is running for them 24.7 in it.
而且这是一种主动式服务，全天候24/7为他们运行。
00:12
We're here at the AI Engineering and New York conference, 20.26
我们现在在AI Engineering and New York conference，2026
00:17
and I'm here with Amman Singh from ARCA and ARCA has recently raised $64 million for
我身边的是来自ARCA的Amman Singh，ARCA最近融资了6400万美元
00:25
an AI native wealth management firm.
要做一家AI原生的财富管理公司。
00:28
So first of all, Amman, can you tell us a bit about your role at
首先，Amman，你能不能跟我们讲讲你在
00:31
ARCA and the financial services you provide and how you're using AI agents to run them?
ARCA以及你们提供的金融服务，还有你们是如何用AI agents来运营这些服务的？
00:38
Yeah, for sure. Thanks for having me.
当然，非常感谢邀请我来。
00:42
My role at ARCA is I lead the AI platform team, so we're building the underlying
我在ARCA的角色是领导AI平台团队，所以我们正在构建底层的
00:46
sort of harness and our capabilities to kind of run the business with AI, which is our ultimate goal.
可以说是一种利用AI来运营业务的手段和能力，这是我们的最终目标。
00:53
In terms of the services that we provide to our clients,
就我们为客户提供的服务而言，
00:56
this is a broad spectrum of financial services.
这涵盖了一系列广泛的金融服务。
00:59
So that includes sort of financial advice, financial planning for their assets
所以这包括财务建议、针对他们资产的财务规划
01:06
and sort of tax planning and insurance and estate planning as people get wealthier.
以及随着人们变得更富有，还有税务规划、保险和遗产规划。
01:13
Maybe you are someone who just sold a business, for example.
比如，也许你是刚卖掉了一家企业的人。
01:17
And there's lots of complexities with all the different partners, etc.
而且涉及所有不同的合作伙伴等等，有很多复杂的情况。
01:20
So we're going to help you think through that.
所以我们会帮助你理清这些。
01:23
We also have clients that are really large technology executives that
我们也有一些客户是大型科技公司的高管，
01:29
are trying to figure out a real plan for retiring early, right?
他们正在试图制定一个真正可行的提前退休计划，对吧？
01:34
And also small families that are trying to bind their first home, for example.
还有一些小家庭，比如正在努力买他们的第一套房。
01:40
So the clients that you have so far, do you have to kind of talk them through using agents
那么到目前为止，你的这些客户，你需要跟他们解释使用agents的过程吗？
01:46
and why you're using agents and tell them a little bit about the technology behind it?
以及为什么要用agents，还要给他们讲讲背后的技术？
01:50
Yeah, so the role of AI at ARCA is not to replace the advisor.
是的，所以AI在ARCA的角色不是取代顾问。
01:56
It is to replace all of the work around the advisor.
它是为了取代围绕顾问的所有工作。
01:59
So our clients don't really interface with AI directly.
所以我们的客户并不直接与AI交互。
02:03
All they experience is a level of service that has not been possible in this industry before,
他们所体验到的是一种这个行业以前不可能实现的服务水平，
02:09
delivered through the empathy and care of a human advisor.
通过人类顾问的同理心和关怀来传递。
02:14
So we're building the machinery behind every client that's constantly thinking about
所以我们正在为每个客户构建背后的机制，不断思考
02:19
what is the market doing and how does it impact Richard?
市场在做什么，以及它如何影响Richard？
02:25
What else is going on in your life?
你生活中还有什么其他事情？
02:27
What are the textures of your life that impact your finances?
你生活中有哪些方面会影响你的财务？
02:31
And we're able to be more proactive, we're more responsive,
我们能够更加主动，更加响应迅速，
02:34
and just deliver a better level of client service using AI that does all of the back office stuff.
并且通过使用AI来处理所有后台工作，提供更高水平的客户服务。
02:41
Okay, so let's get into some of the technical aspects of it.
好的，那我们来聊聊一些技术层面的东西。
02:44
Which models are you using to power your agents?
你们用哪些模型来驱动你们的agents？
02:47
And how do you decide which model is based for which action?
那你怎么决定哪个模型适合哪个动作？
02:51
Yeah, so today we're using the frontier models from the frontier labs like Anthropic and OpenAI.
对，所以今天我们用的是Anthropic和OpenAI这些前沿实验室的前沿模型。
03:01
Claude has gotten our primary model has been but we've gone through several iterations of models
Claude一直是我们主要的模型，但我们已经经历了好几个版本的模型迭代
03:08
over the years. We used to be in Gemini way back when and then OpenAI and then Anthropic.
这些年来。我们以前很早的时候在Gemini，然后去了OpenAI，然后又去了Anthropic。
03:14
And what's guiding all of our decisions are our benchmarks that we have internally for
而指导我们所有决策的，是我们内部为以下方面设定的benchmark：
03:20
all of the various capabilities of our general harness that it has to express in running our business.
我们用来经营业务时必须展现出来的通用框架的各种能力。
03:28
And so just recently, for example, despite the external benchmarks claiming great things for
所以举个例子，就在最近，尽管外部benchmark对以下方面宣称了很好的表现
03:34
Luna, it didn't really quite work out for us as we expected. It was way too slow and it was
Luna，它对我们来说并没有像预期那样真正奏效。它太慢了，而且它
03:40
costing a lot. And so our internal benchmarks are going to guide our model decisions.
成本很高。所以我们的内部基准测试会指导我们的模型决策。
03:45
And as a principle, we want to be model agnostic and sort of not take a dependency on a certain provider
作为一个原则，我们希望保持模型无关，并且不依赖于某个特定的提供商
03:57
or a certain harness from one of these labs and kind of keep that optionality open for us.
或者从这些实验室之一弄个特定的harness，给我们保留这种可选性。
04:01
Okay, so you just mentioned harnesses. So you're using different harnesses depending on which model
好，你刚提到了harnesses。所以你是根据不同的模型来使用不同的harnesses
04:07
you're using. No, we have we have our own kind of built internally that is model agnostic.
你在用的。不，我们有我们自己内部构建的、与模型无关的东西。
04:13
Okay, so when your when a financial advisor is doing something like a taxation report or something
好的，那么当你的——当一个财务顾问在做比如税务报告或者类似的事情时，
04:20
like that, how does the harness decide which model to use? Is that automated?
这个框架怎么决定用哪个模型？那是自动化的吗？
04:25
Well, yes, so we have certain tiers of models based on different complexities of tasks.
嗯，是的，所以我们根据任务的不同复杂程度有不同层级的模型。
04:32
So a task like assisting with a tax plan or doing tax analysis would get our like heaviest tier
所以像协助制定税务计划或者做税务分析这样的任务，会用我们最重的那一层模型。
04:38
of models. Inside of that tier, it is a little bit flexible. So you could use like an open
在那一层里面，是有一定灵活性的。所以你可以用比如一个开源的
04:43
tier model or a sanitary model or a heavy GPT model. But based on there's a lot of like background
tier model或者sanitary model或者heavy GPT model。但基于有很多像背景
04:52
activity that our agents are participating in, right? That is using a lower tier of model based
我们的agents正在参与的活动，对吧？那使用的是较低层级的model
04:58
on the complexity of the task. So that's kind of how we how we make decisions on which model
这取决于任务的复杂程度。所以这就是我们如何决定用哪个模型的方式。
05:03
serves which task. Okay, so is it the financial advisor that's actually making that decision
服务于哪个任务。好吧，那么实际上是财务顾问在做那个决定吗
05:09
while are using the product? Not really. They're not aware of the happening. We've abstracted
在使用产品的时候？并不是。他们并没有意识到正在发生什么。我们把这些都抽象掉了
05:16
this away from them entirely. They kind of just interact with our product as they would
完全远离这些。他们基本上只是像平常一样与我们的产品互动
05:21
and those decisions are made internally by us. Okay, so in your presentation, the
这些决定都是由我们内部做出的。好的，那么在你的演示中，那个
05:28
conference earlier today, you talked about a text bill example. I just wondered if you could maybe
今天早些时候的会议上，你谈到了一个text bill的例子。我只是想知道你能不能也许
05:33
walk us through an example of adjusting a client's portfolio, investment portfolio, you know,
给我们讲一个调整客户投资组合的例子，投资组合，你知道的，
05:40
from identifying a need to maybe rebalance the portfolio to executing trades. Can you talk us
从发现需求，到也许重新平衡投资组合，再到执行交易。你能跟我们聊聊
05:45
through, you know, what parts of that process the agent does and what parts the human advisor does?
通过，你知道，这个过程中哪些部分由agent来做，哪些部分由人类顾问来做？
05:52
Right. So just like in the talk, this is a thing that will span 10 days. And so agents are ultimately
对。所以就像演讲里说的，这是一个会持续10天的事情。所以agents最终是
06:01
responsible for delivering that outcome. But there are certain parts that they do and certain parts
负责交付那个结果。但有些部分是它们做的，有些部分
06:06
that they do not. So in this example, when the client emails us with this whole text bill,
他们其实并没有。所以在这个例子中，当客户给我们发邮件，附上这整段文本账单时，
06:12
there's a piece of the analysis of like, where is this client's money allocated today? What is the
有一部分分析是关于，这个客户的钱今天都分配到哪里去了？那什么是
06:17
most optimal way to liquidate some of that to produce on a thousand dollars? That is best suited for
最理想的方式是把其中一部分变现，来产出在一千美元？那最适合用于
06:23
agents to take on as an analysis task. So they do exactly that. So they have certain limits, though,
让agents把它当作分析任务来承担。所以它们就是这么做的。不过它们有一定的局限性，
06:29
that they can execute a try to correct the amount. So the analysis part is fully automated, but
他们可以执行尝试去修正金额。所以分析部分是全自动的，但是
06:36
they're enabled to actually execute any sort of money movement or liquidation automatically.
他们被授权实际执行任何形式的资金调动或资产变现，而且是自动完成的。
06:41
We have a very deep integration with our custodian that actually gives us APIs to be able to do that.
我们与托管方有非常深度的集成，他们实际上提供了API让我们能够做到这一点。
06:48
But we have strict cardrails around human in the loop for things that would either move money
但我们在human in the loop方面有严格的cardrails，针对那些会调动资金
06:53
or liquidate assets or even directly communicate with our with our end clients, because there's
或变现资产、甚至直接与我们的终端客户沟通的操作，因为我们有
06:58
a level of service that that server service and quality that we expect. So all of the like internal
一定的服务水平和质量预期。所以所有类似内部
07:05
analysis will be done by agents, but the they will also pull in the right person to be able to
分析的工作都会由agents来完成，但它们也会拉入合适的人来
07:10
execute the trade when needed. And they'll give it give the person a plan that they they're
需要时执行交易。他们会给这个人一个计划，他们
07:16
able to review and kind of like adjust. And yeah, go from there. Yeah. So you mentioned that these
能够进行审查，并且有点像调整。然后，是的，从那里继续。是的。所以你提到这些
07:22
processes tend to be, you know, 10 days or at least a longer sort of running time that most
processes tend to be, you know, 10 days or at least a longer sort of running time that most
07:28
sort of agent stuff these days. So when an agent kind of resumes a task maybe days later, then they
现在这些agent之类的东西。所以当一个agent可能在几天后恢复一个任务时，它们
07:36
start it. How do you restore that? How do you check that the plan still fits the clients needs?
开始吧。你怎么恢复那个？你怎么检查这个计划是否仍然符合客户的需求？
07:42
Yeah. So this is an interesting, interesting one to think about is we have tons of these signals
对。所以这个想起来挺有意思的，就是我们有很多这些信号
07:49
like I described in the talk. Some are for example, you can imagine there's an email thread
就像我在演讲中描述的那样。有些是，比如说，你可以想象有一个邮件线程
07:54
that is receiving responses and there's chatter happening. You can imagine that any additional
那是接收响应，还有杂音在发生。你可以想象任何额外的
08:00
response to that goes to the same agent that's kind of like reading that email thread and reacting
对此的回应会发给同一个agent，有点像读取那封邮件线程并做出反应
08:05
to it. But a fresh signal of like a phone call that was made in the middle, that is a brand
对它来说。但中间打进来的那通电话这种全新信号，那才是一个品牌
08:12
new signal that isn't attached with any other ongoing process. And so there's like a fresh agent
一个不依附于任何其他正在进行的过程的新信号。所以就有点像是一个全新的agent
08:17
that wakes up to figure out what this means, right? And the way they way they understand what is
醒来后去弄清楚这意味着什么，对吧？以及他们理解什么是的方式
08:22
happening right now is our internal memory systems. So we have we have this this persistence of
现在正在发生的是我们的内部记忆系统。所以我们有，我们有这种这种持久性
08:29
for every client. What is happening right now? What are we trying to service? Who owns what?
对每个客户来说。现在正在发生什么？我们试图服务的是什么？谁负责什么？
08:35
What has happened so far? What are we tracking for next? What what needs to happen immediately
到目前为止发生了什么？我们接下来在追踪什么？什么需要立即发生？
08:40
in the next couple of days? What are we tracking for the long term, right? Which individual is
在接下来几天里？我们长期在追踪什么，对吧？哪个人是
08:44
responsible for which action? All of this is being tracked by our memory systems that our agents
负责哪个动作？所有这些都被我们的memory系统追踪着，而我们的agents
08:50
contribute to. But also there is a kind of really sort of dreaming process that reflects on the state
做出贡献。但同时也存在一种真正的、类似于做梦的过程，它会反思当前的状态
08:56
of the world as it relates to Richard the client and is able to kind of like update these memory systems
关于世界如何与客户Richard相关的信息，并且能够类似地更新这些记忆系统
09:01
to persist longer as in kind of tasks. So that agent that received a new signal quickly ramps up on
为了在类任务中坚持更久。所以那个接收到新信号的agent会迅速上手
09:08
what the situation is with the client and it'll know okay we're actually executing a big tax bill
客户那边的情况是什么，它就会知道好吧，我们实际上在执行一项大额税单
09:15
and we'll be able to see okay my signal relates to that tax bill. So let me kind of contribute
这样我们就能看到，好吧，我的信号跟那个税单有关。所以让我稍微贡献一下
09:20
to the state of that world. Can you talk a little bit about the safety sort of aspects that you've
到那个世界的状态。你能稍微谈谈你在这方面涉及的安全层面吗？
09:25
maybe built into your product or harness to make sure obviously that no money is lost along the way?
也许内置在你的产品或harness里，确保显然过程中不会损失任何钱？
09:32
Yeah. So all of our safety restrictions are fully deterministic and they take the decision
是的。所以我们所有的安全限制都是完全确定性的，它们来做决定
09:38
making off the agent's hands. So for example certain tools let the agent kind of execute the tool
从agent手里接管过来。所以举个例子，某些工具让agent能够执行这个工具
09:45
itself. So if it's like a simple record keeping kind of tool it will execute it. We also with every
本身。所以如果它像是一个简单的记录保存类工具，它就会执行它。我们还随着每一个
09:51
time a tool is called the rights data. We have a LLM based verification running on that action.
每次调用工具时，就是权限数据。我们在那个动作上跑了一个基于LLM的验证。
09:59
These like fast LLM calls that that analyze every single right or database. So that's
这些像快速的LLM调用，会分析每一个权限或数据库。所以那就是
10:06
another layer of verification we're running. And then we also ingest a lot of like external
我们还在跑另一层验证。然后我们也会摄入大量类似外部的
10:11
information. So there's emails coming in documents coming in phone calls being made which are
信息。所以有邮件进来，文件进来，电话在打，这些是
10:16
targets for prompt injection. So we run sort of verification at the edges that detect presence
prompt injection 的目标。所以我们在边缘处运行某种验证来检测是否存在
10:23
of certain flavors of prompt injection. And then anytime there is a case where data leaves
某些类型的prompt injection。然后任何时候出现数据离开的情况
10:30
our system for example in like a web search in the most obvious case. But also if we're sending
我们的系统，比如说在最明显的网页搜索场景里。但同样，如果我们是在发送
10:36
a calendar invite to Richard right that's an external kind of data where we have both deterministic
给Richard的一个日历邀请，对，那是一种外部数据，我们既有确定性的
10:42
checks on presence of certain information that we know is PII. But also a consensus of small
检查是否存在某些我们已知属于PII的信息。但也要有一个小
10:48
models that ensure that there isn't any kind of data X-Fill being attempted. So we have tons of
确保不存在任何类型的数据X-Fill尝试的模型。所以我们有大量的
10:54
different kind of guardrails built into the harness with the principle that we want to kind of
在harness里内置了不同类型的guardrails，原则是我们想要那种
10:58
contain them at the edges of capability and not earlier than that. Yes. And because you're in
在能力边缘控制它们，而不是更早。是的。而且因为你处于
11:04
the financial industry are there certain kind of regulations that are required that you have to
金融行业是不是有某些必须遵守的监管要求，你必须
11:10
sort of keep to in terms of the technology that you're using or is that is it just more about
可以说是就你使用的技术而言保持这样，还是说这更多是关于
11:15
outcomes and the way you deal with clients. Yeah the industry cares about outcomes but compliance
结果以及你与客户打交道的方式。是的，这个行业关心结果，但合规
11:21
cares about auditability right. And so our system is such that auditability is by default.
很在意可审计性对吧。所以我们的系统是这样的，可审计性是默认的。
11:31
We have every single action that an agent takes produces a change log of what changed
agent 采取的每一个操作都会产生一个变更日志，记录了什么被改变了
11:39
who changed it from which part of the system and when. And this applies not just agents by the way
谁改的、从系统的哪个部分改的、什么时候改的。顺便说一下，这不仅适用于 agent
11:45
it applies to human actors as well. So if you change something in the UI that affects the client
也适用于人类操作者。所以如果你在 UI 里改了某个影响客户的东西
11:51
will log that action. So let's just both get auditability by default but also power some of our
我们会记录那个操作。所以我们既默认获得可审计性，也支撑了我们一些
11:56
agentic capabilities where the same thing that that tracks like a UI change also could be used to
agentic 能力，就是追踪 UI 变更的同一套东西也可以用来
12:03
inform the agent that that happened. So auditability has been a big part of our story both for
告诉agent这件事发生了。所以可审计性一直是我们故事的重要部分，无论是对于
12:10
obviously compliance and regulation but also for our agentic loop. Yes.
显然合规和监管是一方面，但对我们的agentic loop来说也是如此。是的。
12:15
In your presentation I hit a nice line agents prompt people more than people prompt agents.
在你的演讲里我听到一句很妙的话，agent 对人的 prompt 比人对 agent 的 prompt 还多。
12:21
So how do the agents decide who to involve and you know what sort of how much of the context that
那么agents是怎么决定让谁参与进来，以及你知道，大概多少context那些
12:29
they need to provide that person in order to do what they need to do. Yeah. Another interesting
他们需要提供那个人才能完成他们需要做的事情。是的。另一个有趣的
12:35
one. So this is similar to how humans work in our industry is there is a process they were
一。所以这和我们行业中人类的工作方式类似，有一个过程他们当时
12:42
tracking over 10 days. There's multiple actors involved. The actor at step seven may not need to
追踪超过10天。涉及多个行动方。第七步的行动方可能不需要
12:49
know about this just yet right they're pulled in at the right time. And so our agents with our kind
这个现在还不太清楚对吧，它们是在合适的时间被拉进来的。所以我们的agents带着我们的那种
12:55
of like long memory persistence plus the daily reflection on what is happening. Reflect on where
比如长期记忆的持久化，加上对正在发生的事情进行每日反思。反思在哪里
13:02
the state of the world is at the moment and they're able to pull in the right person at that time
世界目前的状态，而且他们能够在那个时候引入合适的人
13:08
when it is when there's a signal that implies that that person must be pulled in. And all of this
当有一个信号暗示那个人必须被拉进来的时候。以及所有这些
13:14
is work that we have done to operationalize the context of how this industry works but not just
这是我们为了把这个行业运作的背景付诸实践所做的工作，但不仅仅是
13:20
that but how does ARCA want it to work and what are our principles of how work happens in this
但那ARCA希望它怎么运作，以及我们对这里工作如何发生的原则是什么
13:25
industry. And that is another advantage of us being sort of a not just a SaaS company. We're
industry。这也是我们的另一个优势，我们不仅仅是一家SaaS公司。我们
13:33
actually not a SaaS company. We are the firm ourselves so we can standardize the way this work
其实不是一家SaaS公司。我们自己就是那家公司，所以我们可以把这项工作标准化
13:38
even happens so that our agents can participate in standard ways that work must happen.
甚至会出现这样的情况：我们的agents能够以标准方式参与，而必须完成的工作也得照常进行。
13:44
So earlier in the conference Anthropic did a workshop about their managed agents infrastructure
所以在会议早些时候，Anthropic做了一个关于他们托管agents基础设施的workshop
13:53
and a lot of it even the orchestration is becoming more automated now. So do you foresee kind of
其中很多甚至编排现在都变得更加自动化了。所以你预见到一种
13:58
using that type of functionality because it obviously presents a danger and that the human
使用那类功能，因为它显然会带来危险，而且人类
14:02
is not necessarily always in the loop anymore. So have you had any thoughts about using that technology
不一定还总是在循环里了。所以你有没有想过用那个技术
14:08
to automate even more of the process. I think we are very bullish on the models
进一步自动化更多流程。我认为我们对这些模型非常看好
14:15
from Anthropic and labs like OpenAI but less so on taking a dependency on a product like
来自Anthropic以及像OpenAI这样的实验室，但在依赖某个产品这方面就没那么强了，比如
14:22
managed agents because managed agents one of their key things is how the model is kind of
因为managed agents，他们的一个关键点就是模型是如何
14:28
persistent and contribute memory as to what is happening in the world and all of the operational
持久的，并贡献关于世界正在发生什么以及所有运营方面的记忆
14:34
laws that they need to be used to be successful. And so letting them own that part of the business
他们需要遵循一些规则才能成功。所以让他们拥有业务的那部分
14:43
is kind of like a risk because owning your intelligence gives you the ability to impact
有点像是一种风险，因为拥有你的智能就赋予了你影响
14:51
the business. If you hand off the intelligence to another provider you lose some of that control
业务的能力。如果你把智能交给另一个提供商，你就会失去一些控制权
14:57
that you need to build something that is unique and tailored to your business to such a great
而你需要这些控制权来构建一些独特的、为你的业务量身定制的东西，定制程度要足够高
15:03
degree that it can actually impact your business. In Anthropic there have also recently released
以至于它真的能影响你的业务。Anthropic 最近也发布了
15:10
financial services so they are partnering with a whole bunch of different firms. Do you think
金融服务，所以他们正在与一大堆不同的公司合作。你觉得
15:15
what they are doing is kind of competitive to what you are trying to do or do you see them as
他们做的事情在某种程度上跟你想做的事情是竞争关系，还是你把他们看作
15:19
different things? Yeah I think Anthropic getting into our kind of industry is very validating
不一样的事情？对，我觉得Anthropic进入我们这个行业是非常有说服力的
15:24
for our industry but the way I think about it is that it is very different to be in a services firm
对我们行业来说，但我的理解是，在一家服务型公司里是非常不同的
15:32
that has access to AI tools versus a firm that was built to be AI native. The firm with AI tools
一家能够使用AI工具的公司，对比一家从创立之初就以AI原生方式构建的公司。那家拥有AI工具的公司
15:39
requires their users to prompt the agents kind of to produce work or to do work
需要用户去prompt这些agent来产出工作或者做工作
15:46
but here like I said in ARCA agents prompt people more than people prompt agents and it is
但在这里，就像我说的，在ARCA里，agent给人们的提示比人们给agent的提示更多，而且它是
15:52
a proactive kind of service that is running for them 24-7 and it's kind of like running the
一种主动式的服务，全天候为他们运行，有点像在运行那个
15:57
business in the most standard way that we expect at the level of service that we aspire to
以我们期望的最标准方式开展业务，达到我们渴望的服务水平
16:04
which looks very different if you are a company with AI tools. Yes. So a little bit more about
如果你是一家拥有AI工具的公司，情况看起来就非常不同。是的。所以再多聊一点关于
16:13
on the security side so can you explain like what in your processes runs inside a sandbox and
在安全方面，你能解释一下你们的流程中哪些部分是在sandbox里运行的吗，还有
16:20
how do you verify like a change to a financial plan before anyone acts on it?
你如何在任何人依据财务计划行动之前验证这个计划的变更？
16:26
Yeah so the way we use sandboxes is that the sandbox is one component of our harness. Our agent
嗯，所以我们使用sandbox的方式是，sandbox是我们harness的一个组件。我们的agent
16:36
isn't actually inside of the sandbox it uses the sandbox to offload computation and we've built
实际上并不在沙盒内部，它利用沙盒来卸载计算，而我们构建了
16:44
systems to make it safe for code inside of that sandbox to access our data. So we kind of like
让沙箱内的代码安全访问我们数据的系统。所以我们有点像
16:52
have this thing which translates the code it wrote into a strict format that we expect
有一个东西，它把自己写的代码翻译成我们期望的严格格式
16:56
throw away the code and you kind of instill guardrails on that format before it reaches your
把代码扔掉，然后你在它到达你的之前，对那个格式设置一些护栏
17:01
back end. So we allow the agents kind of to use code to query our data which means that they can
后端。所以我们允许这些agent用代码来查询我们的数据，这意味着它们可以
17:08
like aggregate it, they can munge it, they can kind of like work with it as they as to see fit
比如聚合它，他们可以随意处理它，他们可以按照自己认为合适的方式去使用它
17:13
to produce such a thing like a like a financial plan. Inside of the same sandbox they produce this
去产出像财务计划这样的东西。在同一个sandbox里面，他们产出这个
17:21
beautiful artifact that we can share with clients that has like it's a PDF of like here's your
一个漂亮的成品，我们可以和客户分享，就像一个PDF，上面写着这是你的
17:26
financial plan and here's how we're planning to kind of like approach your life financially.
财务计划，下面是我们打算怎么从财务角度来规划你的人生的。
17:32
And once that is produced is where we run our verifiers. So it's a certain type of artifact that
一旦生成出来，就是我们运行验证器的地方。所以它是一种特定类型的产物，
17:39
is produced from the sandbox. The moment that is produced we run a bunch of different different
是从沙盒中产生的。产生的那一刻，我们会跑一堆不同的不同的
17:44
verifications on it. Some deterministic that check for presence of like hey did you just put like
对它进行验证。有些是确定性的，检查是否存在类似“嘿，你是不是刚刚放了”这样的内容
17:49
a raw account number in there? You should not do that. So deterministically we can check that
里面有一个原始账号？你不应该那样做。所以我们可以确定性地检查这一点
17:54
but also fast models like like the the spoiler models or Jev to analyze for presence of things
但还有像spoiler models或Jev这样的快速模型，用来分析某些东西是否存在
18:02
that might make this artifact non-compliant. Like for example you really should not be like
这可能会让这个产物不合规。比如说你真的不应该像
18:08
promising returns in that artifact. So that's another thing that we will check before it is
在那个产物中有望获得回报。所以这是我们在它之前会检查的另一件事
18:15
it is kind of publishable. And the last thing I'll say on that is that doing this is actually
这算是可以发表的。关于这一点我要说的最后一件事是，做这件事实际上
18:22
makes our kind of compliance better than if humans were doing it because AI is tireless
让我们的合规性比人类来做更好，因为AI不知疲倦
18:28
at finding even the smallest mention of a non-compliant piece of writing in there. So yeah we'll
在找出其中哪怕最微小的不合规文字提及方面。所以是的，我们会
18:35
continue to invest in more of that in kind of like constrained at the capability edges.
继续在能力边界受限的地方投入更多这类工作。
18:40
Yeah you mentioned Jev there. So you've started to use decision models as part of your process.
对，你刚才提到了Jev。所以你已经把decision models作为你流程的一部分开始使用了。
18:46
So what other products other than Jev are you sort of looking at?
那除了Jev之外，你还在关注哪些其他产品？
18:50
Yeah so what we've noticed is when models are working with financial data right when they're
所以呢，我们注意到的是，当模型在处理金融数据的时候，正好它们在
18:57
answering questions that require financial knowledge they are they're very confident in the knowledge
回答需要金融知识的问题时，他们对自己的知识非常自信
19:04
that they have. So a small example is that models will very happily use the tax brackets
他们拥有的。所以一个小例子是，模型会非常乐意使用税率档次
19:09
from 2025 to come up with an analysis and be like all right that is it but that is obviously
从2025年开始做一个分析然后说好吧就是这样但那显然是
19:16
incorrect and so that is one of the ways in which we're using decision models in our harness
这是不正确的，所以这就是我们在harness中使用decision models的方式之一
19:21
is to check when the model is about to make a financial recommendation or anything any action
就是检查模型什么时候即将做出财务建议或任何行动
19:28
that requires financial rigor will detect and will nudge the model to kind of use our knowledge
这需要财务严谨性，会检测到并会推动模型去使用我们的知识
19:34
base of verified facts about the world that is true today and or use their kind of web search.
一个关于世界的经过验证的事实基础，这些事实在今天是真的，或者使用他们那种网络搜索。
19:39
So it kind of makes it this is kind of a point about the harness and owning the harness that
所以这有点像是一个关于harness以及拥有harness的观点
19:45
lets you do things like this that help your model to be accurate where it matters.
让你能够做这样的事情，帮助你的模型在关键的地方保持准确。
19:51
And do you use decision models and for example recovery if like a trade fails to execute
那你们会用决策模型吗，比如说像交易执行失败时的恢复机制
19:57
is there a decision model involved there or is there another path to get the agent back on track?
那里是否涉及一个决策模型，还是有另一条路径能让agent回到正轨？
20:02
No not today that is not a use case that we're exploring today.
不，不是今天，那不是我们今天在探索的用例。
20:04
Okay so if a step a trade does fail for example how does the agent recover at the moment?
好，那如果某一步交易确实失败了，比如，agent现在是怎么恢复的？
20:14
At the moment we are not doing autonomous trading with agents it is very much a
目前我们并没有用agents进行自主交易，更多是
20:21
the agents will propose that we make trades and it is fully up to kind of our experts to
agents会提出交易建议，然后完全由我们的专家来判断
20:28
figure out if that is the right trade to be making and to execute the trade.
这笔交易是否合适并执行交易。
20:31
Okay so on the same sort of line of that your slides say that every staff edit or cancellation
好的，沿着同样的思路，你的幻灯片上说每一次员工编辑或取消
20:40
comes with a Y and it becomes an evL. So how does that feedback get turned back into the system
都会附带一个Y，然后变成一个evL。那么这个反馈是如何回流到系统中
20:48
as a measurable improvement? Yeah that is another one of the advantages of us being the business
并转化为可衡量的改进的呢？是的，这是我们作为该业务的另一个优势
20:55
and not a SaaS company is that we have the ability to require them to give us feedback.
而不是一家SaaS公司，是因为我们有能力要求他们给我们反馈。
21:02
And yeah the way that we loop the feedback back into improvements is something that you
是的，我们把反馈循环回改进的方式是你
21:06
would expect is all that feedback is piped into a data store that couples that feedback with the
会预期的那样，所有反馈都被管道传输到一个数据存储中，将反馈与
21:15
trace of the agent that produced it. And so we're able to kind of like take that trace
产生它的agent的trace耦合起来。所以我们能够把那个trace
21:22
put it next to the feedback and every night we run this process where we take
放在反馈旁边，每天晚上我们运行这个过程，我们从
21:27
collect traces from the real world from our agents in the wild and run a meta agent on top that
现实世界中收集来自我们agents在野外运行的traces，然后在上面运行一个meta agent，它
21:34
suggests changes to our harness. So maybe tools that were misleading or things that were error or
对我们的harness提出改进建议。所以也许是那些有误导性的工具，或者是出错的东西，或者
21:41
there's missing knowledge in our knowledge base and they'll surface this information to us and
是我们的知识库里有缺失的知识，它们会把这些信息反馈给我们，然后我们就能
21:46
then we'll be able to kind of improve it. Today we still have a human in like a human engineer in
改进它。今天我们仍然有一个人，就是有一个真人工程师在
21:53
the loop before we actually ship a prompt change. I've noticed that agents will make really
loop里，在我们真正上线prompt改动之前。我注意到如果你放任agents自己搞，
21:59
nasty prompt changes if you leave them, leave them to it. Every little fix becomes the whole prompt
它们会做出非常糟糕的prompt改动。每一个小修复都会变成整个prompt，
22:06
and they're unable to stay meta and kind of generalize the fix into a prompt that generalizes.
它们没法保持meta，也没法把修复泛化成一个能通用的prompt。
22:12
So but yeah we have feedback loops from our product that we collect directly as first class
所以但是对，我们从产品中有反馈循环，我们直接作为一等公民来收集
22:18
product features to collect evals into the interim improvements. And you also mentioned in your
将产品功能收集到evals中，融入到阶段性改进里。你还提到了在你的
22:23
presentation about turning knowledge of how our firm works into agent context which is actually
关于把我们对公司运作方式的了解转化为agent context的presentation，这实际上是
22:29
something a lot of forward deployed engineers do try to do as turn in organizations kind of
很多forward deployed engineer确实会尝试去做的一件事，就是在组织里推动某种转变
22:36
internal knowledge into actionable things for agents. What does that look like for your company
把内部知识转化为agent可以执行的东西。这对你们公司来说是什么样的？
22:43
technically and how do you sort of keep it up to date? Yeah so we have a knowledge base of files
技术上你是怎么让它保持更新的？对，所以我们有一个文件的知识库
22:50
that are represented as files and the way we keep that up to date is we've handed it to our experts
这些是以文件形式表示的，我们保持更新的方式是把它交给我们的专家
23:00
to curate an update. So it is a first class feature in our product so on the left hand side there's
来策划一个更新。所以这是我们产品中的一等功能，在左侧有
23:07
a tab called knowledge which has the full operational context that our agents are using.
有一个叫knowledge的标签页，里面有我们的agents正在使用的完整运营上下文。
23:13
And as they work on things they're able to cite knowledge and so it kind of like creates this
而他们在做事情的时候能够引用知识，所以这有点像创造了这种
23:21
this loop of like oh we notice that the knowledge is incorrect and we get surface to us in our
这个循环就像是，哦，我们注意到知识是不正确的，然后它在我们这里浮现出来
23:28
trace analysis and because it is a part of the product itself people are able to go in and
trace analysis，而且因为它是产品本身的一部分，人们能够进去并且
23:34
like edit the skill directly. That's how it kind of works. So you mentioned you're actually
like直接编辑skill。大概就是这样运作的。所以你提到你实际上
23:45
offering these services to clients through you know I presume you have a staff of financial
通过你向客户提供这些服务，你知道我猜你有一支金融方面的员工团队
23:51
advisors in your firm. So can you maybe give us an example of some you know a client that you've
你公司里的顾问。所以你能不能给我们举个例子，比如某个你认识的客户，你曾经
23:58
had that's had a measurable improvement in their investments or taxation or whatever through the
that's had a measurable improvement in their investments or taxation or whatever through the
24:04
agent system and maybe some sort of details about what in the agent architecture kind of helped
agent system，也许还有一些关于 agent 架构中哪些部分起到了帮助的细节
24:10
that client you know get such a good return. Yeah like if you're a recent example this person
你知道的那个客户回报这么好。对，就像如果你是个最近的例子，这个人
24:17
wasn't a client yet but they were they were at the point of becoming a client so we had all of
还不是客户，但他们已经到了快要成为客户的阶段，所以我们有所有
24:24
their documents their statements their current investments etc and this person was already
他们的文件、他们的陈述、他们当前的投资等等，而这个人已经
24:29
working with an advisor and so they meet with us they tell us that hey I have this advisor already
和一个advisor合作，所以他们来找我们，跟我们说嘿我已经有这个advisor了
24:36
they're charging me $1,500 for their services and they're not very responsive to Ian so I'm chatting
他们收我1500美元的服务费，而且对Ian不太上心，所以我在聊
24:45
with you and so our agents have kind of like you know read through all of those documents they've
和你一起，所以我们的agents有点像是，你知道的，把所有这些文档都读了一遍，它们已经
24:50
analyzed them they've analyzed what the person's goals are that they state right they've also
分析了它们，他们分析了这个人所陈述的目标是什么，对吧，他们还
24:56
analyzed the statement of like who the current advisor is right and how much they're charging them
分析了当前顾问是谁、收费多少这样的陈述
25:00
and they discover that there is there are fees that this person was charging then that they weren't
然后他们发现这个人收取了一些费用，而这些费用是他们之前
25:07
aware of and services is it to them and kind of like makes them not only aware of the fact that
不知道的，还有提供给他们的服务，这某种程度上不仅让他们意识到
25:15
they're being overcharged but also aware that that person wasn't even fiduciary so they're
自己被多收费了，还让他们意识到那个人甚至都不是受托人，所以他们
25:21
not obligated to act in their interests so it is this is the sort of like attention to detail
没有义务为他们的利益行事，所以这就是我们的agent在
25:28
that our agents bring about in when they're analyzing a financial picture
分析财务状况时所展现出的那种对细节的关注
25:33
that that sort of would not have been possible if we didn't have have these there's several examples
如果没有这些，那种情况根本不可能实现，这里有几个例子
25:40
of proactivity that our agents kind of like produce of like someone has a deadline coming up
我们的agent所产生的那种主动性，就像有人快到截止日期了
25:46
or they're at risk of missing their RMD or that kind of thing and so this is difficult at scale
或者他们有可能错过RMD之类的，所以这在规模化上很难。
25:53
for humans to keep track of but it is the default behavior of of our agents at ARCA
对人类来说很难追踪，但这是我们ARCA智能体的默认行为
25:58
yeah wonderful well that's probably a good point to stop it so I thank you very much
好，太好了，这大概是个不错的收尾点，非常感谢你
26:03
Amanda for joining us thank you for having me
Amanda 感谢你加入我们，谢谢邀请我