DeepSeek砸510亿做中国版Claude Code,你的AI编程自由要来了DeepSeek invests 51 billion RMB to develop China’s Claude Code, ushering in AI programming freedom


AI free token/AI免费积分

三家顶级AI公司内部数据同时指向同一件事:AI编程Agent不再是程序员的专属工具

━━━━━━━━━━━━━━━━━━

6月25日晚上9点35分,DeepSeek的官方公众号推了一条招聘。标题叫寻找闪亮发光的你。

听起来挺文艺,但内容一点都不文艺——所有部门,规模至少扩大一倍。7个大类,33个岗位,北京杭州同时开招,实习全职全收。

这是DeepSeek成立以来规模最大的一次招聘。而就在十天前,它刚刚完成了510亿元的首轮融资,估值接近4000亿元。

钱到位了,人就成了下一道题。但真正让人心跳加速的,不是510亿这个数字,而是其中一组岗位指向的方向——Agent Harness团队。

说白了,DeepSeek要做中国版Claude Code。

Claude Code是什么,为什么DeepSeek非要自己做

先解释一个概念。Harness,翻译过来叫驾驭系统。你可以把它理解成一辆车的方向盘、油门和刹车——发动机再强,没有这套控制系统,车也开不动。

在AI领域,大模型是发动机,Harness就是方向盘。它决定模型能调用哪些工具、能访问哪些资源、信息怎么在不同子Agent之间流动、什么时候停下来。

Claude Code就是Anthropic做出来的那套方向盘。效果怎么样?一组数据你自己感受:Anthropic内部80%的代码已经由AI编写并提交到主代码库。代码产量提升8倍。

这个数字意味着什么?意味着Anthropic的工程师,现在大部分时间不是在写代码,而是在审代码、定方向。活儿还是那么多,但干法完全变了。

图片

DeepSeek的Harness团队负责人崔添翼,最近每天都在面试,每天都在各个平台发招聘广告。他自己的原话是:作为新成立的部门,目标远大,工作繁重,仍然非常缺人。

这个90后的浙大校友,今年3月才加入DeepSeek。他的团队目前只有三个岗位在招:Harness研究员、工程师和产品经理。对标的就是Claude Code。

用一个不恰当的比喻:大模型是面粉,Harness是面包机。你拿面粉直接吃也行,但有了面包机,才能做出面包来。DeepSeek有了最好的面粉,现在正在造面包机。

三组数据同时指向同一件事

如果只有DeepSeek一家在干这事,那可能只是个公司的战略选择。但三组数据放在一起看,趋势就藏不住了。

第一组数据来自OpenAI。6月26日,OpenAI内部发文披露,他们的编程Agent工具Codex,目前占内部员工输出Token总量的99.8%。去年8月,这个数字还不到10%。不到一年,翻了十倍。

更夸张的是,OpenAI的每位律师和招聘人员,平均85%的产出是在Codex上生成的。法务、财务、招聘部门在今年4月前后集体跨过了使用过半的拐点。

注意,这些不是程序员,是律师和HR。他们不是在用AI写邮件,而是用AI编程工具来完成自动化、数据转换、结构化分析这些以前需要技术人员干的活。

第二组数据来自Anthropic,刚才已经说了——80%的代码由AI编写并提交。

第三组数据来自字节跳动。技术副总裁洪定坤公开分享:过去一年,字节AI代码贡献率增长6倍,AI代码合入率增长超2倍。但同时他也说了实话——主流模型可交付性只有40到60分,结合Harness基建后才能提升至80分。

图片

三组数据,三家公司,不同国家,不同规模,但指向同一件事:AI编程Agent已经不是程序员的工具了,它是全员工具。而Harness,是让这个工具真正能用的关键基建。

为什么Harness这么重要

很多人觉得,大模型能力越来越强,直接用不就行了?

不行。字节的洪定坤把话说得很透:模型生成的东西,和能交付的东西,中间差着一大截。就像你请了个实习生,脑子很聪明,但不知道公司流程,不知道找谁签字,不知道文件该存哪个系统。聪明但不会干活,和能干活的聪明人,是两码事。

Harness就是那套让聪明的AI变成能干活的AI的系统。

它能做什么?举个实际的例子。你对AI说,帮我整理本周销售数据做个报表。没有Harness的AI会给你一段Python代码,然后你自己跑。有Harness的AI会自己打开Excel、读取数据、生成图表、保存到桌面,全程你不用碰键盘。

图片

这就是为什么DeepSeek、Anthropic、OpenAI三家公司几乎同时在押注Harness。模型能力的差距在缩小,但谁能把模型能力变成可交付的成果,谁就赢下一局。

普通人的机会在哪

说完了大公司的棋局,说说跟你我有关的事。

第一,不用等DeepSeek的Harness出来才开始。现在能用的AI编程工具已经不少了。Claude Code、Cursor、Copilot,都可以上手。你不需要会写代码,你只需要能说清楚你想要什么。就像你不会修车,但你会开车就行。

第二,重点不是学编程,是学怎么提需求。OpenAI内部的数据已经证明了:律师用Codex干的事,跟写代码没有半毛钱关系。他们干的是自动化、数据整理、格式转换。你工作中那些重复的、有规律的、需要多步操作的事情,都可以试试让AI编程工具来做。

第三,国产Agent工具会越来越便宜、越来越好用。DeepSeek的定价策略大家都看到了——同样的事,它家的API价格往往是国外模型的三分之一甚至更低。等Harness出来,这个性价比优势还会继续放大。

图片

所以,与其焦虑AI会不会取代你,不如先试试AI能不能帮你干点活。

写在最后

DeepSeek这次扩招有个细节挺有意思。他们还开放了一个叫AI跨界技术人才的岗位,不设专业背景限制,加分项包括不走寻常路、在某个领域做到极致、有创业经历。

梁文锋的逻辑一直很清楚:他不找天才,找有光的人。510亿融资有一个条件——投资人不能挖DeepSeek的人,也不能建议他们出来创业。

一个人守着510亿,把所有赌注押在AGI上。这听起来有点疯狂,但比疯狂更值得关注的是:中国终于有了一家公司,在AI Agent这个最核心的战场上,正面迎战Claude Code了。

这事儿跟你有关。因为当AI编程Agent变成全员工具的时候,谁先上车,谁就先拿到下一张船票。

━━━━━━━━━━━━━━━━━━

 

The internal data from three leading AI companies simultaneously point to one thing: AI programming agents are no longer exclusive tools for programmers.

━━━━━━━━━━━━━━━━━━

At 9:35 pm on June 25th, DeepSeek's official WeChat account posted a recruitment notice. The title was "Looking for you, the shining star."

It sounds poetic, but the content is anything but — all departments are to be expanded by at least double. Seven major categories, 33 positions, recruiting in both Beijing and Hangzhou, open to both interns and full-time candidates.

This is DeepSeek's largest recruitment drive since its inception. Just ten days ago, it completed its first round of financing at 51 billion yuan, with a valuation approaching 400 billion yuan.

With the money in place, people are the next challenge. But what really gets the heartbeat racing isn't the 51 billion yuan figure, but the direction indicated by a specific group of positions — the Agent Harness team.

Put simply, DeepSeek is aiming to create China's version of Claude Code.

What is Claude Code, and why must DeepSeek build its own?

Let's clarify the concept first. "Harness" translates to a control system — think of it as the steering wheel, accelerator, and brakes of a car. No matter how powerful the engine is, without a control system, the car won't move.

In AI, the model is the engine, and the harness is the direction wheel. It determines what tools the model can call, what resources it can access, how information flows between sub-agents, and when to stop.

Claude Code is the control system built by Anthropic. How effective is it? Check out these stats: 80% of Anthropic's code has already been written by AI and committed to the main repository. Code productivity increased by 8 times.

What does this number mean? It means that Anthropic's engineers are now spending most of their time reviewing code and setting direction, rather than writing it. The workload remains, but the way of working has fundamentally changed.

图片

Cui Tianyi, the leader of DeepSeek's Harness team, has been conducting interviews and posting recruitment ads daily. He stated plainly: as a newly established department, the goals are ambitious, the work is heavy, and we are still severely lacking personnel.

This 1990s Zhedu alumnus only joined DeepSeek in March. His team is currently recruiting for three positions: Harness researcher, engineer, and product manager. These are benchmarks for Claude Code.

With an imperfect analogy: the model is the flour, and the harness is the bread maker. You can eat flour raw, but with a bread maker, you can actually make bread. DeepSeek has the best flour, and now they're building the bread maker.

Three sets of data point to the same thing.

If only DeepSeek were working on this, it might be an isolated corporate strategy. But looking at three sets of data together reveals a trend that cannot be concealed.

The first set comes from OpenAI. On June 26th, OpenAI published an internal article disclosing that their programming agent, Codex, accounts for 99.8% of the total token output by internal employees. In August last year, this figure was less than 10%. It has increased tenfold in less than a year.

Even more striking is that 85% of the output from OpenAI's lawyers and recruitment personnel is generated by Codex. Legal, financial, and recruitment departments crossed the 50% usage threshold around April.

Note that these are not programmers; they are lawyers and HR professionals. They are not using AI to write emails but rather to perform tasks like automation, data conversion, and structured analysis that previously required technical staff.

The second set is from Anthropic — 80% of code is written and submitted by AI.

The third set is from Bytedance. Vice President of Technology, Hong Dingkun, shared publicly: AI code contribution rate increased 6x, and AI code merge rate increased over 2x in the past year. However, he also mentioned a critical point — mainstream model deliverability is only 40–60%. It takes integration with a harness architecture to push it above 80%.

图片

Three companies, three datasets, different countries and scales, but all pointing to the same conclusion: AI programming agents are not just programmer tools. They are tools for everyone. And the harness is the key infrastructure that makes these tools truly usable.

Why is the harness so important?

Many people think, "The model keeps getting better. Can't I just use it directly?"

No, you can't. Bytedance's Hong Dingkun put it bluntly: there is a massive gap between what a model generates and what can be delivered. It's like hiring a smart intern who understands the theory but doesn't know the company's workflows, who to sign with, or where to save files. Intelligence is necessary, but knowing how to work is what matters.

The harness is the system that turns an intelligent AI into a working AI.

What can it do? Let's look at a practical example. You tell the AI, "Summarize my sales data for the week and create a report." Without a harness, the AI gives you a Python script, and you run it yourself. With a harness, the AI opens Excel, reads the data, generates a chart, and saves it to your desktop — all without you needing to touch the keyboard.

图片

This is why DeepSeek, Anthropic, and OpenAI are simultaneously betting on the harness. The gap in raw model capabilities is closing, but whoever can turn those capabilities into deliverable results will win.

Opportunities for the average person

After discussing the corporate chess game, let's talk about what this means for you and me.

First, you don't need to wait for DeepSeek's harness to start trying. There are already usable AI programming tools available, like Claude Code, Cursor, and Copilot. You don't need to know how to code. You just need to clearly communicate what you want, like driving a car without needing to repair it.

Second, the key isn't learning programming, but learning how to specify requirements. OpenAI's internal data proves that lawyers are using Codex for things like automation and data cleansing. In your work, any repetitive, rule-based, or multi-step tasks are candidates for AI programming tools.

Third, Chinese-made agent tools will become cheaper and more user-friendly. We've seen DeepSeek's pricing strategy — they often charge a third or less than foreign models for the same capability. Once the harness is available, this cost-performance advantage will widen.

图片

So instead of worrying if AI will replace you, try to find out if AI can help you.

Final thoughts

There is an interesting detail in DeepSeek's recruitment. They have opened a position for "AI Cross-disciplinary Talent," which has no specific professional background requirement. The bonus points are for "thinking outside the box," "achieving excellence in a specific field," and having entrepreneurial experience.

Liang Wenfeng's logic has always been clear: he doesn't look for geniuses, he looks for people with a spark. One condition of the 51 billion yuan investment is that investors cannot poach DeepSeek employees or encourage them to leave.

For a single person to hold 51 billion yuan and bet everything on AGI is a bit crazy. But what is more worth paying attention to is that China finally has a company ready to directly challenge Claude Code in the most critical arena of AI agents.

This matters to you. Because when AI programming agents become tools for everyone, whoever boards the train first gets the next ticket.

━━━━━━━━━━━━━━━━━━

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry