# Forward Deployed Engineering 101
**作者**: vas
**日期**: 2026-05-20T19:10:55.000Z
**来源**: [https://x.com/vasuman/status/2057177266984226892](https://x.com/vasuman/status/2057177266984226892)
---
By the end of this article, you should understand why Anthropic, OpenAI, Google, and other AI companies are looking for FDEs, and how you can capitalize on this demand.
I've done the job myself, hired some of the best in the world to build out our FDE motion at Varick, and noticed there's no real definitive guide to the hottest role in tech right now. This is that guide.
Why do AI companies need FDEs
To become an FDE, the first step is to internalize why AI companies are in desperate need of them.
If you believe intelligence is becoming commoditized, it then follows that the only competitive edge is how and where you use it. In fact, I will go so far as to say that there is no competitive advantage in intelligence alone. Therefore, determining how and where companies use it becomes the most important role and that is the role of a forward deployed engineer.
Businesses hire an Applied AI company (like Varick) that deploys FDEs to help them get the most out of the technology. Doing this gives them access to a team that has already done large-scale AI transformations that make clients move much faster than their competitors, and by doing so, yields massive efficiency gains.
The FDE is a highly skilled engineer who can understand the customer's problems very deeply, write code into a code base they've potentially never seen before, and communicate the business impact to a non-technical decision maker to close the deal. This is a million-dollar hire.
What the role requires
Being an FDE requires you to be on-site with a customer. Palantir's CTO says that you cannot build products for an environment without actually being in the environment itself. We've seen the same thing internally.
The term FDE actually originated from Palantir, and they took being on-site very seriously. In 2010, they worked with the Special Forces group in Afghanistan. The Special Forces would go on the mission in the day, get feedback, and route it to the FDEs who would ship code during the night.
Being in the environment is as necessary for deploying military software as it is for deploying AI. In order to see real efficiency gains, a company needs to be rebuilt around AI from the ground up. And that is only possible through sitting with the customer and building custom agents that are engineered on company-specific data, with company-specific context.
About the role
In our view, there are three main parts of an Applied AI FDE's job: Audit, Evals, and Deployment. Let's break down each one.
Audit: You're onsite with a client, mapping processes/workflows in different teams within the company. For example: two weeks with rev ops, one week with procurement, and a full month with finance.
With each team you sit with, you learn a few things: what their job looks like, where the bottlenecks are, and where you might create agents that deliver value.
Along with understanding the workflows of each team in the company, an important part of the audit phase is determining what should be automated vs what shouldn't. There's a point where agents can create more problems than they solve.
Here are three general principles you can follow to help you decide.
If a workflow can be distilled into rules but the inputs are different (one input is an email, the next could be a PDF, the next is a scanned image), and the work involves calling tools, put an agent in. If the rules and inputs are both predictable, code is faster and cheaper. If the decision needs pattern recognition and domain expertise, leave it manual.
Your clients aren't going to get good ROI if the agent runs five times a month. Look for lengthy, high-volume automations. There needs to be enough volume to matter.
Don't overuse AI when building your agents. Most automation tasks can be done with a series of tool calls and just one call to an LLM as an orchestrating layer. Too much AI leads to unneeded token costs (which compound at scale) and often a lower-quality output.
The final part of the Audit phase is prototyping. See Agents 101 to learn how to build an agent, and Agents 102 to take that agent from demo to production.
Evals: If a customer is spending millions on an AI deployment, they need to know it's working. To do that, an FDE builds detailed evals.
A good eval doesn't just check if the final answer an agent gives you is correct, but also verifies the AI is thinking like a human would. In order to do that, do two things:
Trace the human's steps and grade the AI on each one: A human doesn't solve problems in one move. It's a multi-step process. Map out those steps and see if the AI is hitting the same checkpoints along the way.
Start small with great examples of the intended outcome, then measure everything against them: If you're building a customer support agent, sit with a human and figure out what the best possible answer to a user's query is. Repeat that a few times over a few tasks. Now you know what "great" looks like and can hold the agents to that standard.
Evals prove value to the customer. While everyone says they want AI in their company, there are still many who are skeptical of whether it works. A good agent evaluation is what an executive needs to trust the agent will provide ROI.
Deployment: Avoid large-scale data migrations. Instead, build APIs over an existing data layer (SharePoint or databases) and place a model on top as an orchestrator to query through it. This saves time and money, and more importantly, saves you from the brutal nightmare of ripping out your existing systems. Our clients have spent millions of dollars and multiple years migrating to their latest ERP. The last thing they want to do is replace it again.
Once all of the above is completed, create an execution environment to test the agent safely. This is a sandbox directly in the company's infra so you can run, test, and debug the agent before hitting production.
When moving to production, start slow. Take a small workflow, get it to work, and then layer on additional capabilities. For example, start with an agent that catches bugs, investigates, and writes a ticket summarizing what it thinks went wrong. If that works, only then give it the capability of writing code and pushing a PR.
Start with the smallest unit of autonomy; only then give it the capability to take action.
That is how you go from an audit to deployment as an FDE. Learning these steps is the entire job in itself.
How to become an FDE in 30 days
There are typically three backgrounds that find the most success as an FDE: Consultants, Product Managers, and Software Engineers. Even if you don't fall into any of these, following the 30-day roadmap at the end of this section will exponentially increase your chances of getting the role. Do these in parallel with applying and interviewing.
Consultants/PMs
As a consultant or PM, you should already have the ability to translate data into ROI. That is half the job. But the biggest roadblock for someone with this background is a lack of engineering experience.
A high-quality portfolio can mitigate this. Pick two of these side projects and go all in:
- A production-ready AI agent that can execute an entire process that you used to do manually in your old job. It should be able to call APIs, log its thinking autonomously, and have a failure harness.
- A RAG pipeline built on top of a dataset (choose a custom dataset for whatever industry you're trying to break into: legal docs, medical records, financial filings, etc.).
- An eval framework you built yourself that scores an agent's outputs across multiple dimensions (correctness, format, cost, latency) for different business processes (procurement, accounts payable, etc.).
- An MCP where you can connect an LLM to legacy software that doesn't currently support AI integration.
Do not outsource your understanding to AI. If you take it step by step, these concepts should be quite understandable. There's a reason why this is titled 30 days, not 30 minutes.
Software Engineers
Arguably the most important part of being an FDE is communication. You need to translate what AI can and can't do into something that makes sense to a non-technical VP. If you can't do that, you can't be an FDE.
SWEs should build projects similar to those mentioned in the consulting/PM section, but explain every single component of what you just built. The tech stack, results, iterations you made, business outcomes. Most importantly, you need to have a reason for building those agents in the first place: what was the pain point you're solving for, and how might this go down in a real client interaction?
30-day outline regardless of role
For something more concrete, follow this 30-day plan that will prepare you for almost anything:
Checkpoint 1 (7 days in):
- What is an agent and how an agent loop works: read Anthropic's Building Effective Agents, then write a script that runs the loop: prompt → model → response → next step.
- How to make an agent call a tool: add two tool calls (an API call and a web search) using the Anthropic/OpenAI tool use tutorials.
- How to build proper guardrails: add input validation, a max-step limit, and output filtering before anything reaches the user.
- When to use context window vs external memory: default to context unless state needs to persist longer than the run.
- What is an audit trail and how to build one: log every prompt, tool call, and response with timestamps. This helps find and flag agent errors.
Checkpoint 2 (14 days in):
- How to enforce structured outputs: always return JSON. Read through OpenAI's developer page.
- How to take a demo to prod and what usually breaks: Read Agents 102.
- How to checkpoint: save agent state every n steps to a file so it can restart from the last checkpoint.
Checkpoint 3 (21 days in):
- How retry logic and exponential backoff work: every external call needs retries. On failure, wait 1s, 2s, 4s, 8s, cap at 16s.
- How to optimize cost when deploying agents: three things: cheaper models for cheap subtasks (Opus should be used only for reasoning), cache common prompts, cap max tokens. Track cost per query.
- How to build a golden dataset for evals: start with 20 real queries, label the perfect output yourself. Anthropic's "Demystifying evals for AI agents" covers everything.
- How multi-agent pipelines and parallel architectures work: split the job when one agent can't handle it. One plans, others execute, one synthesizes.
Checkpoint 4 (Final week):
Review the above and communicate everything out loud. Tie everything you can to business metrics.
TLDR
The FDE is the most in-demand role in tech right now. Every company needs AI, but nobody knows how to deploy it.
The job has three phases (audit, evals, deployment). Your job is to understand each phase and its purpose.
Your portfolio and your ability to speak about it are the deciding factors. Build agents, RAG pipelines, eval frameworks, MCPs, etc, and, most importantly, be able to confidently articulate the business use case behind everything you're building.
Lack of communication ability is a deal-breaker for the FDE role. If you can't explain what AI can and can't do to a non-technical decision maker, there won't be a deployment.
Know when AI isn't the answer; this builds trust with the client and more importantly, ROI for the agents in production.
Do these steps and your chances of breaking in will be exponentially higher.
If you want to make the jump to the fastest-growing Applied AI company in SF, Varick is hiring. We're building out the most elite FDE team in Silicon Valley, led by a former COO of Citadel Securities. Apply directly at https://www.varickagents.com/careers.
## 相关链接
- [vas](https://x.com/vasuman)
- [@vasuman](https://x.com/vasuman)
- [32K](https://x.com/vasuman/status/2057177266984226892/analytics)
- [Varick](https://www.varickagents.com/)
- [Agents 101](https://x.com/vasuman/status/2011923440769659132)
- [Agents 102](https://x.com/vasuman/status/2013742184772550851)
- [Building Effective Agents](https://www.anthropic.com/research/building-effective-agents)
- [Anthropic](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview)
- [OpenAI](https://developers.openai.com/api/docs/guides/tools)
- [developer page](https://developers.openai.com/api/docs/guides/structured-outputs)
- [Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents)
- [https://www.varickagents.com/careers](https://www.varickagents.com/careers)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [3:10 AM · May 21, 2026](https://x.com/vasuman/status/2057177266984226892)
- [32.3K Views](https://x.com/vasuman/status/2057177266984226892/analytics)
- [View quotes](https://x.com/vasuman/status/2057177266984226892/quotes)
---
*导出时间: 2026/5/21 10:15:03*
---
## 中文翻译
# 前线部署工程入门
**作者**: vas
**日期**: 2026-05-20T19:10:55.000Z
**来源**: [https://x.com/vasuman/status/2057177266984226892](https://x.com/vasuman/status/2057177266984226892)
---
读完这篇文章，你应该能理解为什么 Anthropic、OpenAI、Google 和其他 AI 公司都在寻找 FDE（前线部署工程师），以及你如何能利用这一需求。
我自己亲自做过这份工作，在 Varick 招聘过一些世界上最好的人来构建我们的 FDE 流程，并且我注意到，对于目前科技界最热门的职位，并没有真正权威的指南。这就是那本指南。
为什么 AI 公司需要 FDE
要成为一名 FDE，第一步是内化 AI 公司为何迫切需要他们。
如果你相信 intelligence（智能）正在变得商品化，那么随之而来的就是，唯一的竞争优势在于你如何以及在何处使用它。事实上，我可以更进一步说，仅靠智能本身是没有竞争优势的。因此，确定公司如何以及在哪里使用智能成为了最重要的角色，这就是前线部署工程师的角色。
企业会雇佣一家应用 AI 公司（如 Varick），该公司部署 FDE 来帮助他们充分利用这项技术。这样做使他们能够获得一支已经进行过大规模 AI 转型的团队，这让客户能够比竞争对手更快地行动，并因此产生巨大的效率提升。
FDE 是一名高技能的工程师，他能够非常深入地理解客户的问题，编写代码进入他们可能从未见过的代码库，并向非技术决策者传达业务影响以达成交易。这是一个价值百万美元的招聘。
该职位的要求
成为 FDE 要求你必须与客户在一起现场工作。Palantir 的 CTO 说，如果不实际身处那个环境中，你就无法为该环境构建产品。我们在内部也看到了同样的情况。
FDE 这个词实际上起源于 Palantir，他们对现场工作非常认真。2010 年，他们与阿富汗的特种部队合作。特种部队在白天执行任务，获得反馈，并将其路由给 FDE，后者在夜间编写代码。
身处环境对于部署军事软件和部署 AI 一样必要。为了看到真正的效率提升，公司需要围绕 AI 从零开始重建。而这只有通过与客户坐在一起，构建基于公司特定数据和公司特定上下文定制的代理才有可能实现。
关于该角色
在我们看来，应用 AI FDE 的工作主要有三个部分：审计、评估 和 部署。让我们逐一分解。
**审计**：你在客户现场，正在绘制公司内不同团队的流程/工作流图。例如：与收入运营 团队合作两周，与采购合作一周，与财务合作整整一个月。
与你坐在一起的每个团队，你都会学到一些东西：他们的工作是什么样的，瓶颈在哪里，以及你可能在哪里创建能提供价值的代理。
除了理解公司内每个团队的工作流外，审计阶段的一个重要部分是确定什么应该自动化，什么不应该。代理有一个临界点，即它们制造的问题比解决的问题还多。
这里有三个你可以遵循的一般原则来帮助你做决定。
如果一个工作流可以提炼成规则，但输入是不同的（一个输入是电子邮件，下一个可能是 PDF，再下一个是扫描图像），并且工作涉及调用工具，那就放入一个代理。如果规则和输入都是可预测的，代码更快更便宜。如果决策需要模式识别和领域专业知识，那就保持手动。
如果代理一个月只运行五次，你的客户不会获得良好的 ROI。寻找冗长的、大批量的自动化。需要有足够的体量才重要。
在构建代理时不要过度使用 AI。大多数自动化任务可以通过一系列工具调用和仅一次 LLM 调用作为编排层来完成。过多的 AI 会导致不必要的 Token 成本（这在规模上会复合）并且通常输出质量较低。
审计阶段的最后部分是原型制作。请参阅《Agents 101》以了解如何构建代理，参阅《Agents 102》以将该代理从演示带到生产环境。
**评估**：如果客户在 AI 部署上花费数百万，他们需要知道它是有效的。为此，FDE 会构建详细的评估系统。
一个好的评估不仅检查代理给你的最终答案是否正确，还验证 AI 的思维方式是否像人类一样。为了做到这一点，要做两件事：
追踪人类的步骤并对 AI 的每一步进行评分：人类不会一步解决问题。这是一个多步骤的过程。绘制出这些步骤，看看 AI 是否在途中达到了相同的检查点。
从预期结果的好的示例开始，然后根据它们衡量一切：如果你正在构建一个客户支持代理，与一个人坐下来，弄清楚对用户查询的最好的可能答案是什么。在几个任务上重复几次。现在你知道了“好”是什么样子的，可以让代理达到那个标准。
评估向客户证明价值。虽然每个人都想在公司中使用 AI，但仍有许多人怀疑它是否有效。一个好的代理评估是高管信任代理能提供 ROI 所需要的。
**部署**：避免大规模的数据迁移。相反，在现有数据层（SharePoint 或数据库）之上构建 API，并在顶部放置一个模型作为编排器来查询它。这可以节省时间和金钱，更重要的是，让你免于拆除现有系统的残酷噩梦。我们的客户已经花费数百万美元和多年时间迁移到他们最新的 ERP。他们最不想做的就是再次更换它。
完成上述所有工作后，创建一个执行环境以安全地测试代理。这是直接位于公司基础设施中的沙箱，因此你可以在上线生产环境之前运行、测试和调试代理。
进入生产环境时，慢慢开始。采用一个小的工作流，让它工作，然后分层添加附加功能。例如，从一个可以捕获错误、调查问题并编写工单总结其认为出了什么问题的代理开始。如果可行，只有这样才给它编写代码和推送 PR 的能力。
从最小的自主单位开始；只有这样才给它采取行动的能力。
这就是作为一名 FDE 如何从审计到部署。学习这些步骤本身就是整个工作。
如何在 30 天内成为 FDE
通常有三种背景的人作为 FDE 最容易成功：顾问、产品经理和软件工程师。即使你不属于其中任何一种，遵循本节末尾的 30 天路线图也会指数级增加你获得该职位的机会。在申请和面试的同时并行做这些。
**顾问/PM**
作为顾问或 PM，你应该已经具备将数据转化为 ROI 的能力。这是一半的工作。但对于这种背景的人来说，最大的障碍是缺乏工程经验。
一个高质量的组合可以缓解这种情况。从这些副业项目中选择两个并全力以赴：
- 一个生产就绪的 AI 代理，可以执行你以前在旧工作中手动执行的整个过程。它应该能够调用 API，自主记录其思维，并具有故障安全装置。
- 构建在数据集之上的 RAG 管道（为你试图进入的任何行业选择一个自定义数据集：法律文件、医疗记录、财务备案等）。
- 你自己构建的评估框架，可以从多个维度（正确性、格式、成本、延迟）对代理的输出进行评分，用于不同的业务流程（采购、应付账款等）。
- 一个 MCP，你可以将 LLM 连接到目前不支持 AI 集成的旧软件。
不要把你的理解外包给 AI。如果你一步一步来，这些概念应该是很容易理解的。这就是为什么这标题是 30 天，而不是 30 分钟。
**软件工程师**
可以说，作为 FDE 最重要的是沟通。你需要将 AI 能做什么和不能做什么转化为非技术 VP 能理解的东西。如果你做不到这一点，你就不能成为 FDE。
SWE 应该构建类似于顾问/PM 部分提到的项目，但要解释你刚刚构建的每一个组件。技术栈、结果、你做的迭代、业务结果。最重要的是，你需要有首先构建这些代理的理由：你要解决的痛点是什么，这在真实的客户互动中可能会如何进行？
**无论角色如何的 30 天大纲**
为了更具体，遵循这个 30 天计划，它几乎可以为任何事情做好准备：
**检查点 1（第 7 天）：**
- 什么是代理以及代理循环如何工作：阅读 Anthropic 的《Building Effective Agents》，然后编写一个运行循环的脚本：prompt → model → response → next step。
- 如何让代理调用工具：使用 Anthropic/OpenAI 工具使用教程添加两个工具调用（API 调用和网络搜索）。
- 如何构建适当的护栏：添加输入验证、最大步数限制和输出过滤，然后再让任何内容到达用户。
- 何时使用上下文窗口与外部内存：除非状态需要保持比运行时间更长，否则默认使用上下文。
- 什么是审计跟踪以及如何构建审计跟踪：记录每个提示、工具调用和带有时间戳的响应。这有助于发现和标记代理错误。
**检查点 2（第 14 天）：**
- 如何强制执行结构化输出：始终返回 JSON。阅读 OpenAI 的开发者页面。
- 如何将演示带到生产环境以及通常会发生什么问题：阅读《Agents 102》。
- 如何设置检查点：每隔 n 步将代理状态保存到文件，以便它可以从最后一个检查点重新启动。
**检查点 3（第 21 天）：**
- 重试逻辑和指数退避如何工作：每个外部调用都需要重试。失败时，等待 1s、2s、4s、8s，上限为 16s。
- 如何在部署代理时优化成本：三件事：为廉价的子任务使用更便宜的模型（Opus 应仅用于推理）、缓存常见提示、限制最大 token。跟踪每次查询的成本。
- 如何为评估构建黄金数据集：从 20 个真实查询开始，自己标记完美的输出。Anthropic 的《Demystifying evals for AI agents》涵盖了一切。
- 多代理管道和并行架构如何工作：当一个代理无法处理工作时，拆分工作。一个规划，其他执行，一个综合。
**检查点 4（最后一周）：**
回顾上述内容并大声沟通一切。尽可能将所有内容与业务指标联系起来。
TLDR
FDE 是目前科技界需求量最大的职位。每家公司都需要 AI，但没有人知道如何部署它。
这项工作有三个阶段（审计、评估、部署）。你的工作是理解每个阶段及其目的。
你的组合以及你谈论它的能力是决定性因素。构建代理、RAG 管道、评估框架、MCP 等，最重要的是，能够自信地阐述你所构建的一切背后的商业用例。
缺乏沟通能力是 FDE 职位的破坏者。如果你不能向非技术决策者解释 AI 能做什么和不能做什么，就不会有部署。
知道什么时候 AI 不是答案；这能建立与客户的信任，更重要的是，为生产环境中的代理带来 ROI。
做这些步骤，你进入这个领域的机会将呈指数级增长。
如果你想跳槽到 SF 发展最快的应用 AI 公司，Varick 正在招聘。我们正在由 Citadel Securities 的前 COO 领导，打造硅谷最精英的 FDE 团队。直接在 https://www.varickagents.com/careers 申请。