Liang Wenfeng has come up with a "Longzhong Plan".

In recent days, the hype around DeepSeek's 70 billion yuan financing has yet to subside, and another clue has emerged: it is shifting its focus towards AI Coding.

Recently, DeepSeek has successively posted two new job openings: Product Manager for Agent Harness and R & D Engineer for Agent Harness.

According to the recruitment information, DeepSeek is transforming the capabilities of cutting - edge models into Agent products. All work outside the model is classified under the Harness category. New hires will join the Harness team and participate in the entire R & D process of DeepSeek's desktop - side Agent product.

These few sentences basically outline the boundaries of DeepSeek's new business.

When Chen Deli, a senior researcher at DeepSeek, was recruiting on social platforms, the title directly stated "Come to DeepSeek to build Code Harness from scratch", and mentioned "Benchmark against Claude Code to build DeepSeek Code Harness".

It is understood that DeepSeek is assembling a new Harness team in Beijing, aiming to build its own code Agent from scratch, with the internal working name pointing to "DeepSeek Code".

This direction is right on the hottest battlefield of large - model commercialization.

Claude Code has turned AI Coding into a project - level development tool. Anthropic disclosed in its February 2026 financing announcement that Claude Code's annualized revenue exceeded $2.5 billion, doubling compared to the beginning of 2026; the weekly active user count of Claude Code has also doubled compared to the beginning of this year.

On the other hand, OpenAI launched the Codex desktop App in February this year, bringing the Coding Agent from the command line to the multi - task desktop environment. Ashley Kramer, the corporate vice - president of OpenAI, recently stated that the weekly active users of Codex increased from 3 million to 4 million within 15 days.

Google also upgraded Antigravity to version 2.0 after I/O 2026. Foreign media reported that it has shifted from an IDE to an independent desktop application and added CLI and SDK, directly targeting Claude Code and Codex.

While assembling the Agent team, DeepSeek is also accelerating its market occupation on the model side. Its official price page shows that after the 25% discount for V4 - Pro ends on May 31st, the discounted price will be officially maintained. From the perspective of the industry, it was generally believed that the long - context and high - cost - performance features of V4 are more suitable for Agent applications.

After the news of the 70 billion yuan financing spread, DeepSeek seems like Liu Bei with a "Longzhong Plan". It is recruiting talents on one hand and pushing the front line towards Code Harness on the other.

01

Regarding Code Harness, DeepSeek is first strengthening its talent pool.

A few days ago, some well - known foreign media successively reported that DeepSeek recruited Cui Tianyi, a former engineer at Jane Street, and the purpose of this talent recruitment is "to catch up in the AI Agent and commercialization competition".

Public information shows that Cui Tianyi joined DeepSeek's newly established AI Harness team. At the same time, more positions are being recruited to support the new flagship desktop - side Agent product.

Looking through Cui Tianyi's resume, one will find that his recruitment has a different meaning from traditional AI researcher recruitment.

Cui Tianyi graduated from the Department of Computer Science at Zhejiang University. During his school days, he participated in the ACM - ICPC regional competitions multiple times and won gold medals. After graduation, he worked at Jane Street's Hong Kong and New York offices for 9 years. In 2022, he co - founded TSY Capital.

Jane Street is a global quantitative trading company. Cui Tianyi worked at its Hong Kong and New York offices for nearly 9 years as a software developer and researcher, covering stock and fixed - income businesses. Obviously, his experience covers trading systems and engineering R & D, rather than being a traditional academic - style AI researcher.

Cui Tianyi later co - founded TSY Capital. This company was established in 2022 and is headquartered in Hong Kong, engaging in program trading in the global stock market.

It uses machine - learning models to generate trading signals and then uses a self - developed low - latency trading system in Rust for execution.

In other words, he is a "veteran in quantitative trading". On one hand, it is reminiscent of the main business field of DeepSeek's parent company, Magic Square. On the other hand, it also shows that his path is different from that of traditional model researchers at DeepSeek and many other AI companies.

DeepSeek used to make a name for itself with base models. Cui Tianyi's entry into the Harness Team indicates that DeepSeek Code needs another type of people: they may not be the authors of large - model papers, but they have experience in real - world systems and know what problems may occur when model outputs are implemented.

From Chen Deli's speech, one can also see the change in DeepSeek's recruitment thinking this time.

Recently, he confirmed on social platforms that the company is assembling a new Harness team, with the direction of "benchmarking against Claude Code to build DeepSeek Code Harness". This time, it is not just recruiting model researchers, but also product managers and R & D engineers.

The job description shows that members of the Code Harness team need to participate in the R & D of DeepSeek's desktop - side Agent product and collaborate with the model research team to bring back feedback from real - world tasks to model training.

By the way, let's mention Chen Deli. This DeepSeek researcher is not only one of the core authors of the R1 paper, but also one of the few insiders at DeepSeek who often appears in public.

At the World Internet Conference in Wuzhen last November, he represented DeepSeek in the "Dialogue of the Six Rising Stars in Wuzhen". When talking about the shortcomings of AI, Chen Deli said that AI does not have the same stable intelligence as humans. After model training, the parameters are fixed and cannot self - iterate in the real world like humans.

Amid the hype of a series of recruitment information, another clue is Xu Mingyu.

Recently, the job - hopping of Guo Daya, a former senior researcher at DeepSeek, has revealed a corner of the talent war in the AI industry.

Related reports show that nearly 70 people have been poached from ByteDance's Seed team in the past year, including Xu Mingyu, a member of the long - term research plan of Seed Edge. Currently, he has joined DeepSeek's model structure group and is engaged in R & D related to model structures.

During his time at ByteDance, Xu Mingyu was the first author of the paper DeltaFormer. The paper discusses the expression limitations of the standard Transformer in state - tracking tasks. The main text mentions tasks such as Python code execution, entity tracking, and chess, and attempts to enhance the Transformer's ability to handle state - space problems with DeltaNet.

Xu Mingyu's joining supplements the model structure capabilities behind DeepSeek Code. Cui Tianyi corresponds to product and system engineering, while Xu Mingyu corresponds to the more underlying continuous - task capabilities. These two lines jointly point to the Code Agent that DeepSeek is working on.

After the Code Agent enters the real repository, the most easily exposed problems are often not the inability to write a function, but forgetting the initial goal after a dozen rounds of modifications, inconsistent logic after cross - file changes, and the inability to continue fixing based on the old state after a test failure. These all test the model's ability to save and update states in continuous steps.

On the other hand, DeepSeek's Harness recruitment has clearly stated its benchmarking targets. The official recruitment information reveals that the Agent Harness position requires candidates to have in - depth experience using products such as Claude Code, Cowork, Codex, Cursor, OpenCode, GitHub Copilot, Manus, OpenClaw, and Hermes.

Obviously, DeepSeek Code needs two types of experience on its growth path: one comes from real - world systems to handle the execution consequences after model output, and the other comes from model structures to handle state maintenance in multi - round tasks. The continuous reinforcement of talent also reveals a new round of the AI talent offensive and defensive battle.

02

A few weeks ago, Guo Daya, a core figure at DeepSeek, switched to ByteDance, which was regarded by the outside world as a significant loss for DeepSeek.

His resume is closely related to the cornerstones of DeepSeek's base models. Public reports show that Guo Daya focused on code intelligence and large - language - model inference and participated in projects such as DeepSeek - Coder, DeepSeekMath, DeepSeek - Prover, V3, and R1. In the DeepSeek - Coder paper, he was listed as the first author.

Around March this year, it was reported that Guo Daya left DeepSeek and then joined ByteDance's Seed. Some public reports show that he is optimistic about the Agent direction, while the priority of Agent at DeepSeek was not high at that time. After joining ByteDance, Seed is initiating organizational integration around Agent and Coding.

Looking back at the present, a few weeks later, DeepSeek, armed with tens of billions in financing, is starting to catch up in the Harness Agent area on one hand and trying to introduce more "Guo Dayas" to enrich its talent pool on the other.

Of course, this talent offensive and defensive battle had already started before Guo Daya's departure, and DeepSeek has never stopped introducing talents in vertical model directions such as Coding.

In the recently announced list of the 2026 ByteDance Scholarships, there is a name worthy of attention.

According to ByteDance, the ByteDance Scholarship has supported 67 young researchers so far. The 2023 recipients, Lu Cheng and Zhu Qihao, are important research contributors to OpenAI's Sora2 and DeepSeek's GRPO respectively.

Among them, Zhu Qihao is a 2024 doctoral graduate from the School of Computer Science at Peking University, focusing on deep code learning. In the DeepSeek team, based on his doctoral thesis work, he led the development of DeepSeek - Coder - V1. Later, he also appeared in the papers related to DeepSeek - Coder - V2 and DeepSeekMath/GRPO and was an early participant in DeepSeek's code model and reinforcement - learning route.

It is worth noting that there is a clear change in Zhu Qihao's position in the iterative process of DeepSeek - Coder.

In the author list of DeepSeek - Coder V1, Guo Daya was listed before Zhu Qihao. By DeepSeek - Coder V2, Zhu Qihao's name was before Guo Daya's.

The order of authors does not directly equate to the size of internal contributions, but for a continuously iterative technical line, there is no doubt that the signal it conveys is that Zhu Qihao is the absolute core figure in the Coding capabilities of DeepSeek's series of models.

Another aspect that cannot be ignored is multi - modality.

In the Chatbot form, DeepSeek's multi - modality shortcoming is not obvious. Users mainly ask questions in text, and as long as the model performs strongly enough in inference, writing, code, and long - context, it can support most usage scenarios. DeepSeek's strongest public perception in the past year has indeed come from R1, V3, V4, and low - cost APIs.

However, the desktop - side Agent will change this premise. If the code Agent enters the developer's computer, it not only needs to answer questions in the chat box but also understand the IDE interface, terminal error reports, browser documents, and screenshot feedback. In other words, screen - reading ability will change from a "bonus" to a basic ability.

Pan Zizheng is a name that has been publicly disclosed more often in DeepSeek's multi - modality business.

In February last year, many media reported on the social media content of Yu Zhiding, a senior research scientist at NVIDIA. Yu Zhiding recalled that Pan Zizheng had an internship at NVIDIA in the summer of 2023. At that time, NVIDIA considered offering him a full - time position, but Pan Zizheng "unhesitatingly" chose to return to China and join DeepSeek. At that time, DeepSeek's multi - modality team only had a few people.

Yu Zhiding also mentioned that Pan Zizheng played a "key role" in multiple projects such as DeepSeek - VL2, DeepSeek - V3, and DeepSeek - R1. He wrote that Pan Zizheng's case is a typical example he has seen in recent years: many of the most outstanding talents come from China, and these talents do not necessarily have to achieve success only in American companies. American researchers can also learn a lot from them.

The capability direction of DeepSeek - VL2 is not far from Code Harness. VL2 covers visual question - answering, OCR, document, table, chart understanding, and visual positioning. These capabilities in the desktop - side Agent exactly correspond to screen - reading, error - reading, web - document reading, and interface - element understanding.

However, DeepSeek's multi - modality has lacked a real breakthrough on the front stage recently. Janus - Pro was released in January 2025, and DeepSeek claims that its image - generation ability exceeds that of DALL - E 3 and Stable Diffusion. VL2 remains at the visual - language - model update in December 2024.

This means that DeepSeek already has the technical foundation for screen - reading and visual understanding, but has not yet established a clear brand label for multi - modality like R1 inference and V4 long - context. If DeepSeek Code really wants to develop a desktop - side Agent, multi - modality is likely to be the direction for subsequent talent recruitment.

Recalling Chen Deli's speech at the World Internet Conference in Wuzhen last year, when talking about AI stability, he said that after model training, the parameters are fixed and cannot self - iterate in the real world like humans. The direction he mentioned is to enable AI to have a stable learning algorithm and establish more links with the real world through multi - modality and other means.

Facing the bottlenecks in the development of large models, DeepSeek has a clear understanding within the company. However, looking back at the slow updates and talent loss in the past year, it is obvious that the large - model competition is a long - term battle that tests capital strength and endurance and requires more "ammunition" supply. After the 70 billion yuan financing is in place, not only Code Harness but also DeepSeek may enter a multi - line narrative to quickly roll out products.

03

Just looking at the base models, DeepSeek has entered the first echelon, but it is not in the leading position.

The new - generation DeepSeek - V4 - Pro has narrowed the gap with leading closed - source models in code benchmarks and Agentic tasks. DeepSeek V4 - Pro reached 80.6% on SWE - bench Verified, approaching Claude Opus 4.6.

In addition, DeepSeek V4 became the top - ranked open - source weight model in the Vibe Code Bench, scoring 49.9%, and is also the only open - source weight model that exceeded 40%.

However, a strong model and a strong product are essentially two different things. The commercialization data of Claude Code is a reference that DeepSeek cannot bypass.

After Anthropic disclosed that Claude Code's annualized revenue exceeded $2.5 billion, the valuation logic for AI Coding has changed. It is no longer just a demo for model companies to showcase their capabilities but a product line that can directly contribute to revenue.

The form of Claude Code has also gone beyond the chat box.

It can work in the developer's code repository and can be used through the terminal, IDE, desktop, Slack, and web. When running in the local terminal, it will request authorization before modifying files or running commands. This design has evolved AI Coding into a project - level assistant.

Running neck and neck with Claude Code are OpenAI's Codex and Google Antigravity 2.0. Large companies are pushing AI Coding from editor plugins to independent development environments.

Meanwhile, this is also one of the most competitive areas for domestic models.

The scores of Kimi and Zhipu in the Coding field have recently entered the first echelon