面试问题默认候选人负担得起 Claude Code Max

核心要点：

• 获取领先的 Agentic Coding 工具正逐渐成为一项招聘筛选条件。
• **差距不只关乎金钱。**用于练习的无偿时间同样重要。
• 早期访问权会迅速产生累积效应。当前的补救措施是一个开始，但力度仍然有限。

在向 Agentic Coding 转型的过程中，贸易基础设施公司 Starboard 的工程团队负责人 James Lowman 开始要求候选人谈谈他们最近编写的三个 Claude Code skill。他认为这是评估候选人对 AI Agent 熟练程度的有效方法，并很快将其作为首选面试问题。随后，他却被现实上了一课。

一位正在面试实习岗位的大学生候选人回答“没有”，并解释说，自己负担不起免费工具以外的任何产品。

“这确实让我措手不及，”Lowman 说。

在当今工程领域，获得领先 AI 工具、Token，以及熟练掌握这些工具所需的时间，是保持竞争力的关键。然而，这些资源并不便宜。那么，如果跟上行业发展和获得录用在很大程度上依赖于获得昂贵 AI 的能力，这意味着什么？

一些 工程领导者开始正视这一问题：他们重新评估招聘流程，并试图在衡量候选人 Agentic 能力与避免让访问资源较少的人处于劣势之间取得平衡。与此同时，这种差距并不能仅靠精心设计的招聘流程来解决：谁现在拥有访问权、谁没有，其累积效应可能会像滚雪球一样扩大。

研究数字鸿沟与劳动问题的牛津马丁 AI 治理倡议研究助理 Matthew Sharp 表示：“领导者需要意识到，居家完成的 Agentic 测试，以及诸如‘谈谈你编写过的 skill 或构建过的 Agentic 工作流’之类的问题，都预设了候选人能够访问能力很强的模型。如果你提出这些要求，你测试的不只是能力倾向；你在一定程度上也在测试谁有钱以这种方式练习。这可能会将一种暂时的资源差距，变成永久性的筛选机制。”

多项 研究表明，生成式 AI具有帮助人们提升水平的作用，对经验较少的工作者帮助最大。但 Sharp 指出，这些研究都控制了参与者的访问条件。获得访问权显然是从 AI 中受益的前提，而这本身就引出了新的问题。Sharp 认为，Agent 让访问问题变得更加复杂。

在题为《Agentic Inequality》的四月论文中，Sharp 与合著者提出了 Agent 访问的三个维度：可获得性（你能否使用某个 Agentic 系统？）、质量（它的能力有多强？）和数量（你能够运行多少个 Agent？）。他说，尽管一直以来都存在让某些人更容易领先的不平等因素，但 Agent 引入了“可扩展的委托”。换言之，你不再只是使用一件工具，而是获得了一种能够代表你采取行动的“劳动力”。

“我认为这是一种新的不平等，它受金钱、算力、专业知识和时间的限制，”Sharp 说。

获得领先 Agentic AI 系统的需求，已经在影响工程师的职业决策。例如，候选人在权衡工作邀请时会协商 Token 预算。类似地，Lowman 描述说，一名实习生在返校后接受他提供的兼职工作机会，主要原因是能够继续使用公司的 Claude Code Max 计划。

“[这几乎到了这样的程度，]他说：‘我甚至不在乎你们是否付我薪水。我只是需要这个，这样我才能继续做项目。’”Lowman 说。

GitHub 在 3 月取消 Copilot 学生计划对高级模型的访问权限，并在 6 月为该计划设定每月 200 个 AI Credit 的上限后，用户强烈表达了失望。他们解释说，自己需要访问高级模型才能跟上行业标准，并指出在新的访问限制下实际能完成的工作少得可怜。

“一条提示词，砰的一下，每月 Credit 就全没了，”一位用户在宣布这项变更的帖子下评论道。该帖获得了超过 6,700 个踩和不足 100 个赞。GitHub 未回应置评请求。

受到影响的不只是学生。经验丰富的工程师也感受到压力：他们需要付费购买多个 Agentic 编程工具，并利用个人时间构建项目，以打动老板和面试官。

工程经理 James Rowe 最近处于待业状态，并正在面试下一份工作。他表示，自己参加的每一场面试都被要求展示 Agentic Engineering 能力。他每月在各种 Agentic 编程工具上花费 100 美元。

“我发现，这是跟上行业发展的唯一方式，”他说。

飞轮效应（而且不是好的那种）

7 月，The Washington Post 的一篇报道因其中一个具体轶事而走红：报道中的一位旧金山工程经理据称要求妻子承担几乎全部照顾年幼女儿的责任，因为他需要专注于成为“AI 原生”人才。报道描述，他在夜晚和周末把自己关在办公室里做项目，最终成功成为“公司里 AI 使用最领先的用户之一”。

这篇报道及其引发的愤怒反应，揭示了 Agent 不平等在工程领域呈现的另一个常被忽视的方面：时间。并非每个人都能把自己的责任转交给别人，从而投入 Agentic 副项目；然而，行业内的讨论通常暗示每个人都应在自己的时间里试验和构建 AI Agent。这种预期也延伸到了招聘中，既体现在“谈谈你最近编写的三个 Claude Code skill”之类的问题中，也体现在招聘流程中需要投入大量时间的环节上。

Starboard 已不再采用居家测试，现在要求候选人与团队进行全天试做，可以线下进行，也可以远程进行。Lowman 表示，这让公司能够轻松地为面试者提供相同的 Agentic 工具访问权，但并非每个人都能参加全天试做。对于这类“特殊情况”，公司会改用居家作业。

不过，居家作业同样存在时间门槛。为了确保候选人在面试过程中都能使用领先的 Agentic 工具，Stack Overflow 平台工程总监 Peter O’Connor 一直在考虑各种方法，例如让候选人访问团队的 AI 工具，或支付一个月的订阅费用。

然而，他发现，解决候选人之间的时间不平等要困难得多。尽管他认为居家作业有助于评估候选人，但他担心其中所需的时间。

“居家作业对于初级工程师，或目前失业、拥有充裕时间的人来说效果很好。可是，一位才华出众的工程师，既是单亲家长，又需要全职工作养家，该怎么办？”他说，“如果我们强制要求在工作时间之外投入大量时间，是否就仅仅因为他们负担不起这段时间，而淘汰了一整类非常出色的人才？”

创造公平竞争环境

为了开始缩小差距，Sharp 建议工程领导者询问候选人：他们在获得 Agentic 工具时是否曾面临不利条件。此外，企业应让所有候选人都能访问同一种工具，以完成居家作业或其他将被评估的项目。

不过，这些都只是小步骤。Sharp 指出，在互联网和手机等此前的数字技术发展过程中，各国政府和国际组织曾投入巨大努力，追踪缺乏访问权的问题，并“弥合数字鸿沟”。

Sharp 认为，政府和研究人员现在需要开始衡量 Agent 不平等，包括追踪不同收入群体和地区的部署差异。他还建议思考如何追踪免费与付费 Agent 层级之间的质量差距，并探索公共 Agent 方案或补贴访问机制，尤其是在教育领域。

Sharp 认为，其中最关键的一点是，这种不平等可能产生飞轮效应。行业现在采取的行动，或未能采取的行动，可能会加剧这种鸿沟，并进一步限制谁能够竞争、获得录用和把握机会。

“当你拥有早期优势时，它会反馈到自身并不断累积。我认为在 Agent 领域，这一点尤其明显，”他说，“那些有能力负担 Token、及早进行试验的人，不仅能完成更多工作；他们还能培养更好地委托任务的技能，这使他们更容易被雇用，也让他们获得更多访问权。”

You have 1 article left to read this month before you need to register a free LeadDev.com account.

Estimated reading time: 7 minutes

Key takeaways:

• Access to leading agentic coding tools is becoming a hiring filter.
• The divide isn’t only about money. Unpaid time to practice matters just as much.
• Early access compounds fast. Current fixes are a start, but they’re small.

Amid the shift to agentic coding, James Lowman, engineering team lead at trade infrastructure company Starboard, started asking candidates to tell him about the last three Claude Code skills they wrote. It seemed like an effective way to evaluate their fluency with AI agents, and it quickly became his go-to question. Then, he got a reality check.

One candidate, a college student interviewing for an internship, replied “none,” explaining they couldn’t afford anything beyond the free tools.

“It really did catch me off guard,” says Lowman.

In today’s engineering landscape, access to leading AI tools and tokens – and time to get good at using them – is key to staying competitive. They also don’t come cheap. So what does it mean if keeping up and getting hired depends heavily on access to expensive AI?

Some engineering leaders are beginning to grapple with this question, evaluating their hiring processes and working to strike a balance around how to gauge candidates’ agentic proficiency without putting those with less access at a disadvantage. At the same time, the gap won’t simply be fixed with a well-designed hiring process: the compounding effect of who has access now and who doesn’t can have snowballing implications.

“Leaders need to realize that take-home agentic tests and questions where they ask ‘tell me about the skills you’ve written or the agentic workflows you’ve built’ presuppose access to very capable models,” says Matthew Sharp, a research associate at the Oxford Martin AI Governance Initiative who studies digital divides and labor. “If you’re asking for those things, you’re not just testing aptitude; you’re also partly testing who can afford to practice in this way. That risks turning a kind of temporary resource gap into a permanent sorting mechanism.”

Several studies suggest generative AI has a leveling-up effect, helping less experienced workers the most. The thing is – the studies all controlled for participants’ access, notes Sharp. Access is an obvious prerequisite to reaping the benefits of AI, which opens up questions on its own. Sharp argues agents complicate the issue of access even further.

In an April paper titled “Agentic Inequality,” Sharp and his co-authors lay out three dimensions of agent access: availability (is an agentic system available to you?), quality (how capable is it?), and quantity (how many agents can you run?). While there have always been inequities that help some people get ahead, agents introduce “scalable delegation,” he says. Put differently, you’re no longer just using a tool, but rather getting access to a kind of workforce that can act on your behalf.

“I think it’s a new inequality, and it’s gated by money, compute, expertise, and time,” says Sharp.

The need for access to leading agentic AI systems is already shaping how engineers make career decisions. Candidates are negotiating token budgets when weighing job offers, for example. Similarly, Lowman describes how continued access to the company’s Claude Code Max Plan was the driving reason an intern took him up on his offer to stay on to do part-time work after he returned to school.

“[It was] almost to the point where he was like, ‘I don’t even care if you’re paying me. I just need that so I can still work on projects,’” says Lowman.

When GitHub removed access to premium models from its Copilot student plan in March, and later instituted a 200 monthly AI credit cap for the plan in June, users loudly voiced their disappointment, explained they need access to premium models to stay on par with industry standards, and pointed out how little they can actually achieve with the newly limited access.

“One prompt and boom all the monthly credits are gone,” commented one user on the post announcing the change, which garnered over 6,700 downvotes and less than 100 upvotes. GitHub did not respond to a request for comment.

Students aren’t the only ones impacted, with experienced engineers also feeling the pressure to pay for multiple agentic coding tools and use their personal time to build projects to impress their bosses and interviewers.

James Rowe, an engineering manager who was recently between jobs and interviewing for his next opportunity, says he was expected to show proficiency in agentic engineering in every single interview he did. He’s been spending $100 per month on various agentic coding tools.

“I find that is the only way to stay current in the industry,” he says.

New York • September 15 & 16, 2026

The pace of change keeps accelerating. See what other leaders are doing about it, at LDX3 New York.

The flywheel effect (not the good kind)

In July, a story in The Washington Post went viral over a specific anecdote about an engineering manager in San Francisco who, according to the report, asked his wife to take over nearly all parenting responsibilities for their young daughter because he needed to focus on becoming “AI native.” It describes how he spent nights and weekends locked in his office working on projects, which allowed him to eventually succeed in becoming “a top user of AI at his company.”

The story and the outraged reaction to it capture another, often overlooked aspect of how agentic inequality is showing up in engineering: time. Not everyone can just transfer their responsibilities onto someone else to dedicate themselves to agentic side projects, yet conversations throughout the industry commonly suggest everyone should be experimenting and building with AI agents in their own time. This expectation extends into hiring as well, both with questions like “tell me about the last three Claude Code skills you wrote” and steps in the hiring process that require significant time commitments.

Starboard has moved away from take-home tests and now asks candidates to do a full-day trial with the team, either in person or remotely. This has made it easy to give interviewees access to the same agentic tools, but not everyone can do a full-day trial, Lowman says. In those “special circumstances,” they’ll opt for a take-home assignment instead.

Take-homes still pose a time barrier, too, however. To make sure candidates are all able to use leading agentic tools during the interview process, Peter O’Connor, director of platform engineering at Stack Overflow, has been considering ways to give them access to the team’s AI tools or cover a one-month subscription.

However, he’s found solving for time inequalities among candidates much less straightforward. While he sees a lot of benefits of take-home assignments in terms of his ability to evaluate candidates, he worries about the time required.

“A take-home assignment works beautifully for a junior engineer or someone who is currently unemployed and has the luxury of time. What about a brilliant engineer who is a single parent working a full-time job to support their family?” he says. “If we mandate a massive time commitment outside of hours, have we just eliminated a whole subset of incredible talent simply because they couldn’t afford the time?”

Level the playing field

To start evening the playing field, Sharp suggests engineering leaders ask candidates questions around if they’ve had any disadvantages in accessing agentic tools. Additionally, they should give all candidates access to the same tool to complete take-homes or any other projects they’ll be assessed on.

At the same time, these are only small steps. With previous digital technologies like the internet and mobile phones, Sharp notes there were massive efforts from national governments and international organizations to track lack of access and “close the digital divide.”

Sharp believes governments and researchers need to start working now to measure agentic inequality, including tracking how deployment is different across income groups and geographies. He also suggests thinking about how to track the quality gap between free and paid tiers of agents, as well as exploring public options for agents or subsidized access, especially in education.

The most crucial aspect of this, according to Sharp, is that the inequality can have a flywheel effect. The actions the industry takes or fails to take now risk exasperating the divide and further limiting who can compete, get hired, and take hold of opportunities.

“When you have an early advantage, this feeds back on itself and kind of compounds. I think with agents, it’s especially sharp,” he says. “So the people who can afford the tokens to experiment early don’t just get more work done; they also build the skills to delegate better, which makes them more hirable, which earns them more access.”