How Forward Deployed Engineering is done at Decagon — Sunny Rekhi

频道: AI Engineer
视频: https://www.youtube.com/watch?v=7wu2hsRfvV0
原文语言: en
统计: 共 20 轮 · Sunny Rekhi 18

[0:01]

[music] [applause]

（音乐）（掌声）

[0:15] Sunny Rekhi

Coming guys, can you hear me just fine? All good? Okay, awesome. Just so I can contextualize this talk a little bit, can I get a show of hands of who here is an engineer or is a forward and in a forward deployed motion at all? Okay. Okay, so I'm in my minority. Okay, awesome. Uh sounds good. So yes, um I'm Sunny. I'm the uh CTO of Forward Deployed Engineering here at Decagon. And today I'll talk about what it is that we do, why we have a forward deployed motion, how it has changed over time as we've gone from 50 people to 500 people over the course of a year. Um how it changes if you're working with a Fortune 20 versus a more mid-market brand. Uh thank thank you all for coming. I hope it's useful. And um yeah, let's get started. So, Okay. Uh okay, so just to give context on what Decagon is. Um for those of you unfamiliar, Decagon is a 24/7 AI customer service agent. So, we've all had the experience of calling into your favorite brand and being told to press one for billing, press two for membership options, etc.

来了来了，各位听得清吗？都还行吧？好，太好了。为了让我大概摸清这场分享该往哪个方向讲，能不能举个手——在座有谁是工程师，或者本身就在做 forward deployed 这类活儿的？好……行，看来我这类人是少数派。好，没问题。那开始吧。我叫 Sunny，在 Decagon 负责 Forward Deployed Engineering，是这块的 CTO。今天我想讲讲我们到底在做什么、为什么我们需要一支 forward deployed 的队伍，以及这一年里公司从 50 人涨到 500 人的过程中，这套打法是怎么变的。还有，服务一家《财富》前 20 强的巨头，跟服务一个中型品牌，做法上有什么不一样。谢谢大家来听，希望对你们有用。那我们开始。先交代一下 Decagon 是干什么的。可能有人不熟悉：Decagon 是一个 7×24 小时在线的 AI 客服 agent。我们都有过那种体验吧——打电话给某个品牌，听筒里传来「账单问题请按 1，会员选项请按 2」……

[1:21] Sunny Rekhi

Or you email into your brand because you need urgent support and you hear back in two or three business days. Decagon replaces all of that. So, instead you pick up and you call your brand of choice and you get a human-like agent who is helping you. You email in, you get a human-like reply right away. So, that's what Decagon does in a nutshell. Um multilingual, omni-channel, etc. Um and then importantly, and again I you know I I I I say this to to help contextualize what our forward deployed motion does. But, you know, we land in our customers to help them with the kinds of complex support workflows that today have to go to humans. Um but once we are there and our agent is learning about the customers and has a relationship with the customers, then we also work with our customers to figure out, "Hey, how can we actually make you more money?" So, one example, you'll see this in the bottom of the slide here, but Hertz, you know, we're all familiar with. Hertz came to us because they had these kind of complex inbound support workflows that indeed it to be offloaded it to an agent.

或者你有急事发邮件过去求助，结果两三个工作日之后才收到回复。Decagon 就是来取代这一整套东西的。你打电话过去，接你的是一个像真人一样的 agent，直接帮你解决问题；你发邮件过去，立刻就能收到一封像人写的回复。这就是 Decagon 干的事，一句话说完。多语言、全渠道，这些都有。还有一点很重要——我说这个是为了帮大家理解我们的 forward deployed 团队到底在做什么——我们进到客户那里，一开始是帮他们接住那些今天还必须转人工的复杂客服流程。但等我们落地之后，agent 开始了解这些终端用户、跟他们建立起关系，我们就会跟客户一起琢磨：「那怎么才能顺便帮你多赚点钱？」举个例子，这页幻灯片底部就有——Hertz，大家都知道这家租车公司。他们最初来找我们，是因为有一堆复杂的入站客服流程确实需要卸给 agent 来接。

[2:26] Sunny Rekhi

But once we were there, uh it turns out, "Hey, we already have like these integrations to your back-end systems. What other communications are you doing with customers?" And one that Decagon now does for them is to reach out proactively to a customer when it's time to renew their car lease or extend it or whatever, and they can do that from within Decagon. So, it is you land We typically we land and help them deflect these sort of inbound support cases, and then we expand into, you know, how do we make you more money? Um Now, we work cross-vertical. We also have really large enterprises, more mid-market brands. These are a subset of what I was approved to talk about. Uh there were way more I wanted to add in there, but our our marketing had got mad at me. We have, you know, the the top left we have our financial institutions, the bottom right we have our, you know, your favorite tech brand.

可我们进去之后就发现：「诶，我们已经打通了你们后台系统的这些集成了，那你们还有哪些跟客户的沟通场景？」于是 Decagon 现在还帮他们做一件事：在客户的租车合约该续约、该延期的时候主动去联系客户——这些都能在 Decagon 里完成。所以路径是先落地、先帮他们把入站的客服工单分流掉，然后再往「怎么帮你多赚钱」这个方向扩。我们的客户跨了很多行业，既有超大型企业，也有中型品牌。这页上放的只是我被批准可以公开讲的一部分，我本来还想多放几个，但市场同事已经对我发火了。左上角是我们的金融机构客户，右下角是大家熟悉的科技品牌。

[3:18] Sunny Rekhi

Uh and this is relevant because, as I'll talk about briefly, uh the kind of forward deployment you have to do is vastly different based on both the size of the enterprise and also the vertical. Okay. So, I imagine this is the case for a lot of agentic companies, but Decagon has effectively two kinds of forward deployment engineering. Number one is taking that AI customer service agentic brain and making it work for your enterprise. So, the same way that you train a human, you give it instructions on what to do when a user asks X, how you respond back to it, what sort of brand tonality you have, what actions do you take on behalf of the user. All of this like configuring of that human of the of that [clears throat] agent brain is one form of our forward deployment motion, where we work with the customer, we figure out what does success look like for you, how do you want the agent to speak, what sort of user intents do you actually want the agent to hand off to a human instead.

这一点之所以重要，是因为——待会儿我会简单讲——你要做的 forward deployment 具体长什么样，会因为企业的体量不同、所处行业不同而天差地别。好。我猜很多做 agent 的公司都是这样，Decagon 的 forward deployment engineering 实际上分成两类。第一类，是把这个 AI 客服 agent 的「脑子」拿过来，让它在你这家企业里跑得通。就跟你培训一个新员工一样：你要告诉它，用户问 X 的时候该怎么办、该怎么回，品牌的语气调性是什么样的，你可以代表用户执行哪些操作。所有这些配置 agent 大脑的活儿，是我们 forward deployment 的一种形态——我们和客户一起，搞清楚「对你来说什么才算成功」「你希望 agent 用什么口吻说话」「哪些用户意图你其实是希望 agent 直接 handoff 给人工的」。

[4:20] Sunny Rekhi

That's the left half of this diagram. And we have a team, which I'll talk about briefly, who is like really good at configuring the agent. Largely, this can also happen within the UI. And then on the right side is um the the previous speaker alluded to this as well. Forward deployment forward deployment engineers are the front line for customer product asks. And it is their job to figure out, "Hey, enterprise A made this ask. I know in 2 weeks enterprise B is also going to have the same ask." And this happens with stunning regularity. So, I want to make sure when I solve enterprise A's problem, I'm solving it for B, C, D, and E before they've even had a chance to express it. So, there's the two kinds of forward deployment engineer that we have internally, configuring the agent and then making sure all the problems that you interact with the enterprise that that come up in that in in in in the context of conversation also get brought back into the product.

这是这张图的左半边。我们有一支团队专门干这个，特别擅长配置 agent，待会儿我会讲讲。这些活儿大部分也能在产品 UI 里完成。而右半边——上一位讲者其实也提到了——forward deployed engineer 是承接客户产品诉求的第一线。他们的任务是判断：「A 企业提了这个诉求，我知道再过两周 B 企业也一定会提同一个诉求。」而这种事发生的频率高到让人吃惊。所以我要确保：当我解决 A 企业的问题时，我是在为 B、C、D、E 一起解决——赶在他们开口之前。所以我们内部的 forward deployed engineer 就是这两类：一类配置 agent，另一类负责把你跟企业打交道过程中、在一次次对话里冒出来的所有问题，反哺回产品里。

[5:16] Sunny Rekhi

Which brings me to a really important point. In fact, it's so important I wish I had a slide for it, but uh at Decagon, forward deployment engineering is identical to product engineering. Uh it's the same bar, it's the same reporting structure, uh often like the same team, because the the the the delineation between what is historically forward deployment versus product engineering is super super blurred now. Uh when I'm speaking with a Fortune 20 and they express a pain point, that is often a product feature that needs to get built and prioritized. And so that line between I'm a forward deployed person and I'm a person who works on the product um is gone. Uh it's the same it's the same person and and that's represented in our in our in our org chart. So, um early on, I mean, Degagon is is an example of sort of canonical hypergrowth. A year ago we were at 50 people.

这就引出了一个特别重要的点。事实上它重要到我真希望自己给它单独做了一页幻灯片——在 Decagon，forward deployment engineering 和产品工程是完全等同的。一样的招聘标准、一样的汇报线，很多时候就是同一支团队。因为历史上那条区分 forward deployment 和产品工程的界线，如今已经模糊到几乎不存在了。当我在跟一家《财富》前 20 强的客户聊、他们说出一个痛点时，那往往就是一个需要被造出来、需要被排进优先级的产品功能。所以「我是 forward deployed 的人」和「我是做产品的人」之间那条线已经没了。就是同一批人，这一点在我们的组织架构图上也是这么体现的。说回早期——Decagon 算是那种典型的超高速增长样本。一年前我们还只有 50 个人。

[6:14] Sunny Rekhi

[snorts]

（轻笑）

[6:14] Sunny Rekhi

Now we're at 500. And uh the scale is not slowing down. So, actually shameless plug, if you are interested in uh a new role, sunny@degagon.ai is my email. Please let me know. I'll make sure your your resume {slash} profile gets front of my people. Anyway, back to the back to the talk. Um so, historically we had agent software engineers and they did it all. They did that configuring of that agent brain sitting side by side with our customer. Uh this is again things like what is the tonality of the agent, what sort of voice do you want it to have, both literally the voice, but also the the way it speaks. Um how do I integrate it into your back-end systems so that it could take action on behalf of the users? This can be something simple like I want to reset my password, so the agent needs to have back-end access into your, you know, authentication system. And it can be something sort of far more complex than that.

现在是 500 人。而且这个增速还没有慢下来的意思。所以厚着脸皮打个招聘广告：如果你在看新机会，我的邮箱是 sunny@decagon.ai，欢迎联系我，我保证把你的简历/资料直接送到该看的人手上。好，回到正题。历史上我们只有一个岗位叫 agent software engineer，什么都干。他们坐在客户旁边，一起配置 agent 的那个「脑子」。还是那些事：agent 的语气调性是什么、你希望它用什么样的声音——既是字面意义上的音色，也包括它说话的方式。以及怎么把它接进你的后台系统，让它能代表用户执行操作。这可以很简单，比如「我要重置密码」，那 agent 就需要能访问你的认证系统后台；也可以复杂得多。

[7:07] Sunny Rekhi

And they also did some of that like platform work like customer A has this feature request and then building that back into the product. Now that we're 500 people, uh we start thinking a lot more about how do we design the Degagon system so that it can scale. And effectively we we we broke apart this agent software engineering role into two specialized lanes. One is the agent builder and these are like Degagon pros. They have a lot of intuition for the various models that power our platform. How do you make them work for the use case that the enterprise requires? Um largely living within the UI to the extent possible, flagging when things need to go off UI and how do we bring that into the product. And then secondly, we have agent software engineers. Again, these are the front-line enterprise makes product request. Making sure that gets incorporated back into the product. And um, this is I think like a super super uh, important insight, uh, which is there is routinely this temptation of Okay, customer A made this request and they're so important to us and they want it done ASAP and maybe I'll just go prompt Codex and Cloud Code to just do it for me.

同时他们还得干一部分平台的活儿——比如 A 客户提了这个功能需求，然后把它做回产品里。现在我们到了 500 人，就得开始认真想：Decagon 这套体系该怎么设计才扛得住规模。于是我们实际上把原来那个 agent software engineer 的角色拆成了两条专门的赛道。一条是 agent builder，这些人是不折不扣的 Decagon 高手，对驱动我们平台的各个模型有很强的直觉，知道怎么把它们调教到符合某家企业的具体用例。他们尽可能地在 UI 里完成工作，同时负责标出哪些事情 UI 里做不了、以及怎么把这些需求带回产品。另一条是 agent software engineer，还是那句话——他们是企业提出产品需求时的第一线，负责确保这些需求被吸收回产品。还有一点我觉得极其重要：你会反复面对一种诱惑——「A 客户提了这个需求，这客户对我们太重要了，他们希望马上就要，那我干脆丢给 Codex 或者 Claude Code，让它们直接给我写出来算了。」

[8:15] Sunny Rekhi

But the scarce skill now that AI coding is so good, the scarce skill is actually exercising restraint. Uh, and saying, you know, really thinking about how does this going to scale to sort of future customers? And part of this is our ethos. Like we we build agents to be owned by the customer and so if it turns into a black box of like prompts and patches and that's not good for us or them. It's far too brittle. Um, but also, uh, when you're a forward deployed person, this is this is kind of uh, this is incumbent upon you to to be exercising this restraint of like, let me not do the easy one-off thing, but rather make sure whatever I am building is architected in a way that future customers benefit from. So this will come up in in the remainder of my 10 minutes here, which is uh, always thinking about how do I make this one ask benefit the remainder of the customers? Um, so I I I put this slide here not to sort of toot our own horn, but to actually talk about what it looks like to achieve success. Uh, and in in our case, we've learned like early on when you're scoping the deal, like literally when the very first conversations, you want to figure out ahead of time what does success look like for the customer. And really narrowing that down, ideally getting it in writing so that there is like no miscommunication along the way.

但现在 AI 写代码这么强，稀缺的技能反而是「克制」。你得真的去想：这套东西能不能撑到未来的客户身上？这背后也是我们的价值观——我们做的 agent 是要交给客户自己拥有的，如果它变成一堆 prompt 加补丁拼出来的黑盒，那对我们、对客户都不是好事，太脆了。而且当你是做 forward deployed 的人，克制本来就是你分内的事：别图省事做一次性的活儿，而是保证你做的任何东西，在架构上都能让后面的客户跟着受益。接下来这十分钟我讲的内容都绕不开这一条——怎么让眼前这一个需求，惠及其余所有客户。我放这页 slide 不是想自吹，而是想聊聊「成功」到底长什么样。以我们的经验，很早——真的是在谈单子、第一次对话的时候——你就要提前搞清楚，对这个客户来说成功是什么样子。而且要把它收得足够窄，最好落成白纸黑字，这样整个过程中不会出现理解偏差。

[9:38] Sunny Rekhi

Like and when I say what does success look like, I mean what are the metrics you're trying to hit, what sort of channel that you want support on, maybe that's a phone call, maybe that's email, maybe that's text, maybe it's WhatsApp, whatever. But really narrowing like what is your pain point, what is the ideal outcome you want, and then we can race to go build that out. Um but I think again, back to sort of lessons for forward deployed folks, uh there especially when you're dealing with a large company, there's a temptation to just get started. And uh and this is partly a reflection of how AI coding has changed engineering generally, but now there's a lot of effort that has to go up front in requirements gathering, making sure you're aligned on what actually has to get built uh before before going to do it. Uh this has been a really good learning for us. So, uh we try now, given that we have like a a ton of customers across various verticals, we have found it's really helpful to have industry experts that get staffed, the same kind of deal. So, if I am working on financial service A, B, and C, when financial service D company comes around, ideally I have a core core group of folks who have experience with those customers uh working with this new logo.

我说的「成功长什么样」，具体是指：你要冲的指标是什么？你想在哪个渠道上做支持——可能是电话，可能是邮件，可能是短信，可能是 WhatsApp，都行。关键是把它收窄：你的痛点是什么，你想要的理想结果是什么，然后我们就可以全速去把它做出来。再回到给 forward deployed 这帮人的经验——尤其面对大公司的时候，人是很容易忍不住直接开干的。这在一定程度上也反映了 AI 写代码对工程这行的整体改变：现在前期得花大量力气在需求梳理上，先把「到底要造什么」对齐了再动手。这一点我们学到了很多。所以现在，因为我们的客户横跨很多行业，我们发现配「行业专家」非常管用，做法也是一样的：如果我做过金融服务的 A、B、C 三家，那等第四家金融公司进来的时候，最理想的是让这批有相关客户经验的核心成员去接这个新 logo。

[10:51] Sunny Rekhi

And the idea here is like a lot of that knowledge compounds. Like A, you can like speak in the lingo of this customer, and therefore there's a lot more credibility there. There's a lot more There's a lot like a much faster ramp up. And [snorts] uh a lot of the agent building, sort of the way you think about success carries over. So, uh this has been very helpful for us, and and ultimately it's all about, you know, uh making every deployment uh faster than the last one. Um I mentioned earlier that as a forward deployed and and by the way, I say forward deployed engineering, but really it's just like all forms of forward deployment. I mentioned earlier that one of the big things you have to do is to always think about how do I solve this customer problem in a way that extends to other customers. The other thing that I think is always helpful to keep top of mind is how do I make it so that I'm empowering the rest of the business to solve this problem? And this is specifically if you're an engineering.

这么做的道理是，很多知识是会复利的。第一，你能用客户的行话跟他对话，可信度立刻就不一样，上手也快得多。而且 agent 怎么搭、你对「成功」的理解方式，这些经验都能迁移过去。所以这对我们帮助非常大，说到底就是让每一次交付都比上一次更快。我前面提到，作为 forward deployed——顺便说一句，我嘴上说的是 forward deployed engineering，但其实各种形式的 forward deployment 都算——我前面说过，你必须一直想的一件大事是：怎么把这个客户的问题，解成一个能延展到其他客户身上的方案。另一件我觉得也该时刻挂在心上的事是：我怎么做，才能让公司里其他人也有能力解掉这个问题？这一条尤其是说给做工程的人听的。

[11:46] Sunny Rekhi

So, for example, uh Deckagon's ethos is you should be able to configure this agent completely via natural language. And so, if you ever have an engineer needing to do something that needs to get upstreamed back into the product. Uh and so, this is sort of a funnel that we have of like, "Look, Deck front forward play engineering, they're the front line for customer asks." Um but really it should get it should get sort of scaled across the business. And one example of this, and I mentioned it later as well, is like, let's take an integration. Let's say Deckagon needs to integrate into like some some CRM. Uh early on in our history, we were actually like building custom integrations time and time again. And then we thought enough is enough. After like the 25th one, we're like, "I don't know how many more are coming up." Uh so, let's just like build it a self-serve way. And now what took an engineer custom code writing can now be self-served by the customer or built by our agent building team. So, it's all about how do you scale the work that you're doing.

举个例子，Decagon 的信条是：这个 agent 应该完全靠自然语言就能配置出来。所以一旦出现某件事非得工程师亲自动手，那它就该被反推回产品里去。我们内部有这么一个漏斗：Decagon 的 forward deployed engineering 是客户需求的第一线，但这些需求最终应该在整个公司层面被规模化掉。举个例子——后面我还会再提——比如集成。假设 Decagon 要接进某个 CRM。早期我们真的是一个一个手写定制集成，写到第 25 个的时候我们受不了了，心想「鬼知道后面还有多少」，那干脆做成自助的吧。于是原本要工程师写定制代码的活儿，现在客户自己就能配，或者由我们的 agent 搭建团队来做。核心就是：你怎么把自己手上的活儿规模化。

[12:41] Sunny Rekhi

Um also very relevant, depending on the kind of forward deployment work you do, is especially in the enterprise, wanting to prove value as fast as possible. For those of you who work especially in the Fortune 500, you're going to get hit with the entire What's the expression? Kitchen sink or the entire kitchen? Something like this. Uh but the idea is how do you prove value as fast as possible? So, in our case, Deckagon can become arbitrarily complex. You can support all sorts of channels, all sorts of very complex user intents. We try to figure out how do we demonstrate value ASAP and not have like a multi-month deal or sorry, multi-month uh time to prove value. And once we're there and we're adding value, then we expand, right? Because ultimately all of our customers are a multi-year partnership. And so, we want to make sure we're we're helping you across your entire support flow and in your revenue generating workflows, but it's important as a forward deployed person to figure out, how do I prove value right away and build your build your motion around that.

还有一点也很关键，具体看你做的是哪类 forward deployment——尤其在企业客户这边，你要尽可能快地证明价值。在座如果是服务 Fortune 500 的，客户会一股脑把所有东西都砸过来，那个说法叫什么来着？kitchen sink，还是整个厨房都端过来？大概这个意思。总之关键是：你怎么用最快的速度证明价值。拿我们来说，Decagon 可以做得任意复杂——支持各种渠道、各种非常复杂的用户意图。但我们会去想，怎么最快把价值跑出来，别搞成一个几个月才见效的单子，不对，是几个月才见到价值的周期。等价值跑出来了、我们确实在产生价值了，再往外扩，对吧？因为归根到底，我们跟客户都是多年的合作关系，我们希望覆盖你整个支持流程，也覆盖那些带来营收的工作流。但作为 forward deployed 的人，重要的是想清楚：我怎么马上把价值证明出来，并围绕这一点组织自己的打法。

[13:44] Sunny Rekhi

Um So, customers will often come to you come to folks and say, uh I want you to do XYZ. And and often they're right. But, I think as a forward deployed engineer or forward deployed person of any sort, you're you should treat yourself as an advisor rather than just an executor, right? You're you're both. So, um you're also on the front line of, "Hey, how do I make AI work for the enterprises?" And you have so much knowledge because you're seeing it repeated across every single customer. And so, what we do at Akkio gone is we actually ingest your historical support data uh and we tell customers that hey, like if you automate this first or this first, this is where actually you'll see the highest ROI. Um and sometimes that's not actually what the customer had reached out about. Uh and I imagine there's analogs to this across all sorts of verticals, but it's important to keep in mind that your job isn't just an executor. It is of course to be an executor, but it is also to be an advisor. Uh and to not underrate the fact that you have this domain expertise by being forward deployed across many companies so that you have this knowledge base that's really valuable for the customer to tap into.

客户经常上来就说：我要你做这个这个那个。而且往往他们是对的。但我觉得，作为 forward deployed engineer，或者任何形式的 forward deployed 角色，你应该把自己当成顾问，而不只是执行者——其实两者都是。你同时也站在「怎么让 AI 在企业里真正跑起来」的第一线，而且你手上的信息量特别大，因为同一件事你在每一个客户身上都见过。所以我们在 Decagon 的做法是：把你历史上的客服数据接进来，然后告诉客户，如果你先自动化这一块、或者先动那一块，ROI 是最高的。有时候这跟客户当初找上门想解的问题并不一样。我猜各行各业都有类似的对应。总之要记住：你的工作不只是执行——当然执行也是本职——但同时也是当顾问。也别低估这件事：正因为你在很多家公司做 forward deployed，你手里攒下的是一个对客户非常有价值、可以随时来取用的知识库。

[15:03] Sunny Rekhi

Every time at Akkio gone, someone has to do something manually, we try to make sure it gets upstream back into the product. So, I mentioned the integration earlier, but this is I think a good mental model for folks to have if you're on the front lines. How do we smoothen out that path? Custom becomes self-serve. Custom becomes self-serve. Um this has become like a guiding ethos for us uh and I suspect it is the case across every kind of forward deployment motion. So, I'd encourage everyone in this audience to to um to to keep this top of mind. Like, okay, I'm doing this I'm doing this one-off thing. Presumably other people in the company also are bringing it back into the product. Okay. So, um Decky on is really interesting in that it was started by I think now they're in their early 30s, but it was I think at the time they're in their early 20s. Oh, sorry, late 20s. Um and so, what did what did we do right uh to to sort of deserve the place that we have?

在 Decagon，只要有人不得不手工做某件事，我们就会想办法把它反推回产品里。前面说的集成就是一个例子。我觉得这对站在一线的人来说是个很好的心智模型：怎么把这条路铺平？让定制的东西变成自助的。定制变自助。这已经成了我们的一条准则，我怀疑在任何形式的 forward deployment 里都成立。所以我建议在座各位都把这条记在心上：好，我现在在做这件一次性的活儿，那公司里大概率还有别人也在做，那就把它做回产品里去。好。Decagon 有意思的地方在于，创始人现在大概三十出头，当年创办的时候我记得才二十出头——不好意思，是快三十。那我们做对了什么，才配得上今天这个位置？

[16:00] Sunny Rekhi

And I think one is we're known in the industry to move really really fast on customer asks. And part of this is just like it's a very hard-working group of folks. Um so, that's like a big reason that we got here uh is that we just move really fast deal by deal. Thing number two, we've earned trust with customers that we are advisors, not just executors. So, we'll we'll we'll we'll we'll be able to tell you based on what we're seeing across all the customers, based on the data you give us um what is going to be the highest ROI for you. And then number three, we've been really good at um making sure we productize custom work. But, the way we think about this has changed a lot in the last year because again, a a year ago we were 50 people, could all fit on, you know, a lengthy lunch table. And now we're 500. So, now we think a lot about designing the system. So, every time now now we're very rigorous about sharing knowledge across deployments, but uh making sure you extend uh the field the people in the field feed information back to the platform, making sure the agent compounds every single time it interfaces with the customer. So, if the agent interfaces with customer A, you improve that for customer B.

我觉得第一，我们在行业里以「响应客户需求特别特别快」著称。这里面很大一部分就是这帮人真的能拼。所以我们能走到今天，一个大原因就是一单一单地跑得非常快。第二，我们在客户那儿赢得了信任——我们是顾问，不只是执行者。我们能基于在所有客户身上看到的情况、基于你给我们的数据，告诉你哪一块 ROI 最高。第三，我们很擅长把定制工作产品化。但这件事我们的想法在过去一年里变了很多，因为一年前我们才 50 个人，一张长餐桌就能坐得下；现在是 500 人。所以现在我们会花很多心思在「设计这套系统」上。我们对跨交付的知识共享变得非常严格：一定要让一线的人把信息回流到平台上，让 agent 每一次跟客户打交道都在复利。也就是说，agent 服务了客户 A，你要把这份改进带给客户 B。

[17:11] Sunny Rekhi

And it's all about sort of taking knowledge from the field and bringing it back into the product. Uh and so, just to wrap up here, uh sort of the a few of the few of the themes. Number one, obviously you have to make sure you can figure that agent, do whatever the customer wants. But number two, uh make sure that it gets fed back into the product. And three, mind that funnel that I mentioned earlier. You're on the forward, you're on you're on the field, but you want to make sure it scales and make sure it improves uh every every uh future customer interaction. Uh again, my my email is sunny@deciagon.ai. I'll also be out here. Folks have questions. Thank you for coming to talk and I hope this is helpful.

说到底，就是把一线的知识带回产品里。最后收个尾，几个主题：第一，你当然得能把那个 agent 配出来，客户要什么就配成什么。第二，一定要让这些东西反哺回产品。第三，留意我前面说的那个漏斗——你人在前线、在一线，但你要保证它能规模化，保证它让未来每一次客户交互都变得更好。再说一遍，我的邮箱是 sunny@decagon.ai，我一会儿也会在外面，大家有问题可以来找我。谢谢大家来听，希望这些对你们有帮助。

[18:06]

[music]

[音乐]