文 | 超聚焦
DeepSeek补票Agent时，曾经的候选人郭达雅已换乘字节。
10月16日，DeepSeek成立产品部，将Harness、Web、App、Platform四个团队整合到一起，由拿下过六枚ACM亚洲区域赛金牌但出身量化崔添翼负责，统筹产品、设计和研发。
这家曾经以模型研究闻名的大模型公司，开始把更多精力放到Agent和产品体系上。模型之外，如何让AI真正执行任务、稳定运行，正在成为新的竞争焦点。
就在几个月前，DeepSeek可能错过了最适合驾驶这艘巨轮的船长。
2026年3月，R1核心贡献者之一郭达雅离职，加入字节跳动。离职原因是他看好Agent方向，但该方向在DeepSeek内部优先级不高；而字节跳动愿意投入充足资源，让他能够专注于Agent研究。
如今，DeepSeek开始加码Harness、整合产品团队，而那个曾经想做Agent的人，已经去了另一家公司。
如果郭达雅没有离开，今天站在产品部负责人位置上的会不会是他？DeepSeek又能否驶向今天的方向，更早拿到大结果？
DeepSeek想留的人,字节先给了舞台
郭达雅离开DeepSeek，像一个擦肩而过的故事。
1994年出生的郭达雅，2023年博士毕业后加入DeepSeek。在这家以技术研究著称的公司里，他参与了DeepSeek-Coder、DeepSeek-Math、V3、R1等多个重要项目，是代码智能与大模型推理方向的核心研究人员之一。
尤其是R1背后的强化学习技术，让郭达雅的名字进入了更多人的视野。DeepSeek用相对高效的训练方法，让模型在数学、代码等任务上展现出强大的推理能力，也由此改变了外界对中国大模型研发实力的看法。
但对郭达雅而言，把模型做得更聪明，可能还不是终点，他看好的下一个方向，是Agent。
模型回答问题只是AI能力的一种体现，但未来的AI产品，需要根据目标拆解任务、调用工具、执行操作，再根据结果调整下一步行动。
这意味着，竞争的重点不再只有模型本身，还包括围绕模型搭建的整套执行体系。但公司还没准备好把下一步放到最高优先级。
对于郭达雅来说，这可能正是他与DeepSeek之间出现分歧的地方。他看好Agent，却没能等到这一方向在DeepSeek内部获得更高的优先级。
与其继续等待，他最终决定换一个平台，去做自己认为更重要的事情。
据晚点报道，郭达雅在2025年下半年便开始考虑离职，直到2026年3月才最终离开DeepSeek。期间，阿里、腾讯、字节等公司都曾向他抛出橄榄枝。阿里据称还向他提供过post-training方向负责人的机会。
最终，他选择了字节Seed。
今年4月，澎湃新闻通过多方信源确认郭达雅已经加入字节。相关报道显示，他成为Seed中Agent方向负责人之一，职级为L8。
字节方面虽然否认了外界流传的“亿元年薪挖人”说法，并解释了Seed团队的薪酬与激励机制，但也可见字节对郭达雅的认可程度。
除了客观的薪酬水平，字节能吸引郭达雅的另一个重要原因，或许是公司正在把Agent作为重要的技术方向推进。
从组织架构来看，2026年初，字节启动了针对Agent和Coding的组织整合，字节希望进一步强化智能体与代码智能方向，而郭达雅此前积累的研究经验，恰好与这条路线相契合；而在技术侧，晚点也曾报道，字节下一代模型的重点就放在了优化Agent能力之上。
这意味着，Agent不只是郭达雅个人看好的研究方向，也是字节接下来准备重点投入的技术方向。对他而言，加入字节Seed，意味着自己的志趣所在与公司的研发重点对上。
然而造化弄人，几个月后的DeepSeek，又变成郭达雅曾经最向往的地方。10月，DeepSeek开始加码Harness，将Web、App、Platform等团队整合到一起，成立产品部。曾经被放在模型之外的产品与执行体系，如今被推到了更重要的位置。
只可惜故人已乘黄鹤去。
郭达雅等不到的转身,崔添翼从Harness开始补
如果说郭达雅的故事，是一个提前看好Agent的人与DeepSeek擦肩而过，那么崔添翼的登场，多少都有些出人意料了。
一个做量化交易的，为什么会被梁文锋选中，来负责DeepSeek的Agent体系？
或许要从崔添翼过往经历中寻找答案，他不是传统意义上的产品经理，也不是一路从大模型研究中成长起来的技术负责人。相反，崔添翼曾经在华尔街做了近九年量化交易。
早在算法竞赛时期，他就展现出对抽象问题和系统设计的兴趣。《背包问题九讲》中，他尝试用函数式编程的思路重新审视经典算法问题；进入Jane Street后，他面对的则是另一种更复杂的系统：在真实市场中，再精巧的策略，也必须依赖底层工程架构才能兑现。
这两段经历看起来相距甚远，却有一个共同点：不满足于让单个模块表现出色，而是思考如何把不同模块组织起来，让整个系统有效运转。
而DeepSeek如今要解决的，恰恰也是一个系统性问题。模型研究决定能力的边界，但当公司开始把Harness、Web、App、Platform整合到一起，考验的就不再只是某项技术能否突破，而是能否把不同环节组织成一套完整的产品体系。
崔添翼接手Harness后，交出的第一份成绩单，确实有些出人意料。
8月13日，DeepSeek Harness正式开源，采用MIT协议，核心思路是“一切皆插件”。模型、工具、会话、沙箱乃至执行循环，都可以被替换和组合。它不是只给DeepSeek自己的模型服务，而是试图让开发者能够自由搭建、改造自己的Agent系统。
项目发布后迅速走红。截至10月中旬，仓库已已经积累约25万颗GitHub Star。对于一个面向开发者的底层工程项目来说，表现足够亮眼。
更值得关注的，是它在实际任务中的表现。
近日，基于DeepSeek Harness开发的插件催生了B站评论区的「大肥鱼」现象：一群由AI驱动的账号顶着鲸鱼娘头像，自动搜索视频、读取字幕，甚至以预设人设发表评论，让DeepSeek从代码仓库一路游进了真实的互联网。
在更量化的测试场中，Composio曾用30个真实SaaS任务测试不同的Agent框架。在同一模型和特定工具环境下，DeepSeek Harness完成了20个任务，Claude Code完成了19个；前者平均每个任务消耗约8.9万Token，后者约65万。
崔添翼没有重新训练一个模型，而是试图从模型之外入手，让现有能力以更低的成本、更灵活的方式发挥出来。
但崔添翼的故事，显然还没到庆功的时候。
10月，DeepSeek把Web、App、Platform三个团队也整合进产品部，产品、设计和研发统一由他统筹。这意味着，他需要面对的，不再只是一个开源项目的技术表现，而是DeepSeek整套产品体系如何协同运转。
这与他的量化经历形成了一种宿命般的呼应。在交易系统里，策略本身只是起点，最终结果取决于整套系统能否稳定执行。
DeepSeek的模型能力已经摆在那里，接下来要看它能否被组织成一套真正可用的系统。
Harness只是起点，真正的考验，是能否把它变成一套成熟、稳定、让用户愿意持续使用的产品。
郭达雅曾经看好Agent，选择离开DeepSeek继续押注；崔添翼则从Harness出发，接过了把这条路继续铺下去的任务。
只是，不知道某个夜晚，当崔添翼继续埋头打磨Harness时，会不会也想起那个曾经与DeepSeek擦肩而过的郭达雅。
又或者，郭达雅会不会在字节的某个深夜，想起那头持续深度求索的巨鲸。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Zoo, focus.
When DeepSeek filled Agent, the former candidate, Guo Daya, was replaced by byte.
On 16 October, DeepSeek established the Product Department, which brings together Harness, Web, App and Platform teams, with six ACM Asia regional gold medals, but quantization Choi, in charge of integrated product, design and development.
The company LLM, once known for modelling research, began to focus more on Agent and product systems. Beyond the model, how to make AI truly operational and stable is becoming a new focus of competition.
Just a few months ago, DeepSeek may have missed the most suitable captain to drive the ship.
In March 2026, one of the R1 core contributors, Quadaya, left and joined the byte beat. The reason for the departure was that he was looking in the direction of Agent, but that direction was of low priority within DeepSeek; and byte was willing to devote sufficient resources to allow him to focus on Agent research.
Now DeepSeek has started to add Harness, the integration team, and the man who once wanted to be Agent has gone to another company.
If Gordaia hadn't left, would he have been in charge of the product today? Could DeepSeek be in the direction of today and get the big results sooner?
DeepSeek wants to stay. byte went first to the stage.
Gordaya left DeepSeek like a story to pass by.
Born in 1994, Guadaya, joined DeepSeek after doctorate in 2023. In this company, known for its technical research, he is involved in several important projects, such as DeepSeek-Coder, DeepSeek-Math, V3, R1, and is one of the core researchers in the direction of code intelligence and LLM reasoning.
In particular, the intensive learning technology behind R1 has brought the name of Gordaya to a wider audience. DeepSeek, with its relatively efficient training methods, has allowed models to demonstrate strong reasoning on tasks such as mathematics, code, and so forth, thereby changing external perceptions of China’s LLM research and development power.
But for Gordaia, making the model smarter might not be the end, and the next direction he looked forward to was Agent.
Model answers are only an expression of AI capacity, but future AI products need to adapt next step operations to target dismantling missions, call tools, perform operations and adjust to results.
This means that the focus of competition is no longer only on the model itself, but also includes the entire implementation system built around the model. But the company is not ready to put next step on top priority.
This may be where he and DeepSeek disagree. He looked after Agent, but did not wait for this direction to gain higher priority within DeepSeek.
Instead of waiting, he finally decided to change the platform to do what he considered to be more important.
Later, it was reported that Guadaya had begun to consider leaving his post in the second half of 2025 and did not finally leave DeepSeek until March 2026. During this period, companies such as Ali, Tencent and Z113Z threw olive branches at him. Ali also allegedly provided him with an opportunity to head the post-training direction.
Eventually, he chose byteSeed.
In April this year, news reports confirmed, through multiple sources, that Guo Daya had joined byte. According to relevant reports, he became one of the head of the Agent direction in the Feed, grade L8.
byte, while denying the “billion-year-old excavators” that has been circulated to the outside world and explaining the salary and incentive mechanisms of the Seed team, can also be seen in the extent to which byte has recognized Guadaya.
In addition to the objective level of remuneration, another important reason why byte can attract Guadaya may be that the company is moving forward with Agent as an important technology.
In terms of organizational structure, early in 2026, byte initiated organizational integration for Agent and Coding, and Z113Z wanted to further strengthen the AI agent and code intelligence orientation, which was matched by Guadaya ' s previous research experience; on the technical side, it was also reported later that the focus of Z113Z ' s next generation model was on optimizing Agent ' s capabilities.
This means that Agent is not only the direction of the study that Guo Daya personally sees, but is also the next technical direction for byte to prepare for focused input. For him, joining byteSeed meant that his interest was aligned with the company ' s research and development priorities.
However, a few months later, DeepSeek became the preferred place for Guadaya. In October, DeepSeek began to add Harness, bringing together the Web, App and Platform teams to form the Product Department. Products and delivery systems that were once outside the model have now been pushed to a higher level.
It is a pity that the people have taken the yellow crane.
I can't wait to turn around.
If the story of Guo Diya had been passed by a man who had seen Agent in advance with DeepSeek, the arrival of Choi Tim Wing was somewhat unexpected.
Why is a quantization dealer chosen by Liang Wenfeng to run the Agen system in DeepSeek?
Perhaps it would be appropriate to look for answers from Choi, who is not the traditional product manager or the technical director who grew up from the LLM study. On the contrary, Choi Tim Wing had made quantization deals on Wall Street for nearly nine years.
As early as the Algorithmic Competition, he showed interest in abstract questions and system design. He tried to re-examine the classic algorithm in a functional programming approach in "Speech 9"; when he entered Jane Street, he faced another more complex system: In real markets, sophisticated strategies also have to rely on a bottom-up engineering architecture to deliver.
The two experiences appear to be far apart, but have one common denominator: not content to make individual modules work, but to think about how to organize different modules and make the system work effectively.
And DeepSeek is a systemic problem. Modelling studies determine the boundaries of capacity, but when companies begin to integrate Harness, Web, App, Platform, the test is not just whether a technology can break through, but whether it can organize the different links into a complete product system.
The first report card handed over by Choi Tim Wing after taking over Harness is indeed somewhat unexpected.
On 13 August, DeepSeek Harness officially opened with the MIT agreement, with the central line being “everything plugin”. Models, tools, sessions, sandboxes and even implementation cycles can be replaced and combined. It does not only serve DeepSeek's own model, but rather seeks to free developers to build and adapt their Agent systems.
Project publish was quickly red. By mid-October, the warehouse had accumulated some 250,000 GitHub Stars. For a bottom-up project for developers, performance was sufficiently bright.
What is more noteworthy is its performance in the field.
Recently, a plugin developed on the basis of DeepSeek Harness gave rise to the phenomenon of "fat fish" in the Bilibili comment area: A group of AI-driven accounts with whale Queenheads automatically search for videos, read subtitle, and even comment on pre-sets, allowing DeepSeek to swim from the code warehouse into the real Internet.
In the more quantization test field, Composio used 30 real SaaS missions to test different Agent frames. In the same model and specific tool environment, DeepSeek Harness completed 20 missions and Claude Code completed 19; the former consumed an average of approximately 89,000 Tokens per mission and the latter approximately 650,000.
Rather than retrain a model, Choi Ming Wing is trying to start from outside the model so that existing capabilities can be deployed in a cheaper and more flexible manner.
But the story of Choi Tim Wing is clearly not yet ripe for celebration.
In October, DeepSeek also integrated the Web, App and Platform teams into the Product Department, where product, design and research and development were integrated. This means that what he needs to face is not just the technical performance of an open source project, but how the DeepSeek product package works in tandem.
This was a fateful response to his quantization experience. In the trading system, the strategy itself is the starting point, and the final outcome depends on the stable implementation of the entire system.
DeepSeek's modelling capability is already in place, depending on whether it can be organized into a truly usable system.
Harness is just the starting point, and the real test is whether it can be turned into a mature, stable set of products that users are willing to use on a sustainable basis.
Gertaya had looked after Agent and had chosen to leave DeepSeek to continue the bet; Choi Tim Wing had left Harness to take over the task of continuing the road.
It was just that one night, when Choi Tim Wing continued to grind Harness, he also remembered Guadaya, who had been with DeepSeek.
Or is it possible that Guadalya will think of a giant whale at some late night of byte, which continues to demand in depth?
Auto-fetched and translated by AI Atlas. All rights belong to the original author.