Anthropic 的 1 亿美元赌注：拆解 Claude Frontier Academy

每一家企业都在竞赛，把 AI 塞进自己的组织；但真正能把 AI 在一家真实公司里跑起来的人，成了最难招到的人才。2026 年 10 月 2 日，Anthropic 用 1 亿美元和一份「医学院式」的培养方案回应了这件事：Claude Frontier Academy，目标是在 2027 年底前培养出 1 万名 Frontier Deployed Engineer（FDE，前沿交付工程师）。这不是一门网课，也不是一张考试证书，而是一段带着真实项目、要被反复评估的驻留经历。本文把它的结构、首批学员、提名机制和对整个行业的信号讲清楚。

这个项目刻意模仿医学院的培养路径

Anthropic 宣布了什么

据 Anthropic 官方公告，这是一笔 1 亿美元的承诺，目标是到 2027 年底培养 1 万名 FDE。首批学员来自埃森哲、Bain、Capgemini、澳大利亚联邦银行、Deloitte、麦肯锡、摩根士丹利和诺和诺德等机构；据 CNBC 报道，首批小组就包含这些公司的工程师。项目目前在旧金山、纽约和伦敦开设，并计划在 2027 年底前扩展到 1 万人规模。它建立在 Claude Partner Network 之上——Anthropic 称该网络已有 4.6 万家企业、超过 17.5 万份 Claude 认证，并有近 4000 人完成了 Basecamp 沉浸式课程。

# The program, in one object. All figures are from Anthropic's
# official announcement dated October 2, 2026.

CLAUDE_FRONTIER_ACADEMY = {
    "launched": "2026-10-02",
    "commitment_usd": 100_000_000,
    "goal": "10,000 Frontier Deployed Engineers (FDEs)",
    "deadline": "end of 2027",
    "first_cohorts": [
        "Accenture", "Bain", "Capgemini",
        "Commonwealth Bank of Australia", "Deloitte",
        "McKinsey", "Morgan Stanley", "Novo Nordisk",
    ],
    "locations": ["San Francisco", "New York", "London"],
    "entry": "nomination by your organization",
}

为什么瓶颈是「人」

Anthropic 的判断是：企业里，一小群技能极深的人，驱动了 AI 交付的绝大部分成果——难的是把这批人才「造」出来。它引用自家全球业务发展与伙伴关系负责人 Steve Corfield 的说法：一个具备正确技能、能访问 Claude、又深谙企业如何运转的高主动性小团队，足以改变一整家公司。而据 CNBC 引用的 Draup 数据，今年金融业对「前沿交付工程师」这类岗位的需求已经在飙升。换句话说，模型能力不是当前的瓶颈，把模型变成生产系统的人是。

# The path is a residency, modelled on medical training: learn from
# practitioners, practise on real cases, get assessed before you act
# alone. Two badges, in order.

PATH = [
    ("1. In-person block",
     "multi-day with Anthropic engineers and licensed instructors; a
      simulated enterprise deployment from use case to security review
      to handover; ends with a graded practical"),
    ("2. Badge: Claude Resident Engineer",
     "earned after passing the graded practical"),
    ("3. 12-week residency",
     "lead a real Claude project inside your own organization, with
      Anthropic engineers and your cohort as support"),
    ("4. Badge: Claude Frontier Deployed Engineer",
     "earned after a second assessment; first expected in early 2027"),
]

# Prior experience building AI agents is explicitly NOT required.

学员带着自己公司的真实 Claude 项目进入驻留

驻留项目怎么走

FDE Residency 照搬的是培养医生的路径：医生跟着执业医师学、在真实病例上练、通过评估后才独立执业。学员先参加一个为期数天的线下项目，与 Anthropic 工程师和持证讲师一起，完整走一遍「模拟企业交付」——从选对用例，到安全评审，再到交接——最后用一个全新场景做一次计分实操。通过者获得「Claude 驻留工程师」徽章，进入为期 12 周的驻留期：在自己所在的组织里主导一个真实的 Claude 用例，期间由 Anthropic 工程师支持，并从同批学员中互相学习。12 周结束时再评估一次，通过者获得「Claude Frontier Deployed Engineer」徽章，首批预计在 2027 年初出现。

# Who the program is for. Anthropic lists the traits it screens for
# on nomination; treat this as the profile you would want to develop.

IDEAL_CANDIDATE = {
    "role": "hands-on software engineer",
    "has": [
        "strong fundamentals",
        "a record of building with LLMs",
        "experience helping others adopt AI",
        "a habit of tackling the business's most important problems",
    ],
    "required": "a named Claude project to lead on return",
    "not_required": "prior experience building AI agents",
}

def is_eligible(engineer):
    return engineer.nominated and engineer.has_project

一个驻留项目该长什么样

把模拟交付里那套关口放进你自己的真实项目，就是很好的准备。项目至少要回答五个问题：修的是哪个坏掉的流程、给谁用；Agent 做什么、以及明确不做什么；数据范围、凭据和人工审批点在哪里；上线前用什么带标签的样本集、以什么阈值判定通过；交接时的运行手册、负责人和回滚路径是什么。这五道关口不是官僚流程，它们正是评估时会被追问的地方。提前把它们写进项目文档，等于提前把评估跑了一遍。

# What a residency project needs to survive the assessment. The
# simulated deployment runs use case -> security review -> handover,
# so build the same gates into your real project from day one.

RESIDENCY_GATES = [
    "problem: which broken process does this fix, and for whom",
    "design: what the agent does, and explicitly what it must not do",
    "security: data scope, credentials, human approval points",
    "eval: a labelled set and a pass/fail threshold before rollout",
    "handover: runbook, owners and a rollback path",
]

def ready_to_ship(project):
    return all(g in project.artifacts for g in RESIDENCY_GATES)

首批学员来自埃森哲、麦肯锡、摩根士丹利、诺和诺德等

怎么进去，以及怎么「等价地」得到这些技能

参与方式是提名制：由所在组织提名，企业可以让自己的 Anthropic 客户团队或伙伴客户经理确认资格；项目要求学员带着一个「有名字」的 Claude 项目回来主导。值得强调的是，Anthropic 明确表示并不要求你有构建 AI Agent 的经验——它要的是扎实的软件工程基本功、用 LLM 构建过的记录、帮别人采用 AI 的经验，以及愿意啃业务里最难问题的习惯。即便拿不到席位，右侧清单也能让你在没有驻留的情况下，练出同样的能力。

# A pragmatic prep checklist if you want your team nominated, or
# simply want the same skills without waiting for a cohort slot.

PREP = [
    "pick one high-value, well-scoped process to automate",
    "write the use case in the language of the business, not the model",
    "define the eval before you write the prompt",
    "get a security review early, not at the end",
    "measure time saved and quality, not 'AI usage'",
    "package the result as a reusable internal pattern",
]

# Anthropic's own framing: a small team of high-agency people with the
# right skills can transform an entire company.

这件事意味着什么

Anthropic 在可能上市的前夕，把增长押在了「人」这一层：它不是在比拼基准分数或 token 价格，而是在建设一条经过认证、能把 Claude 部署进大型企业的工程师管道。这背后是一个更普遍的趋势——当模型越来越像基础设施，竞争会转向谁能把它真正落到业务里。对工程师个人而言，信号也很清楚：把「能用一个模型做出 demo」升级成「能把一个模型带进生产、被评估、被治理」，才是接下来几年最值钱的那部分能力。