Insiders: DeepSeek is forming a Harness team to compete with Claude Code

Author｜Wang Bo, Jiazi Light Year

"Jiazi Light Year" has learned from sources close to DeepSeek that the company is organizing a new Harness team focused on code intelligent agent products, internally benchmarking against Claude Code from Anthropic.

DeepSeek's senior researcher Chen Deli recently confirmed this on social media, stating, "DeepSeek is organizing a new Harness team to work on products and research in the Harness direction," and bluntly said, "In simple terms, it is benchmarking Claude Code to create DeepSeek Code Harness."

This is not an ordinary recruitment.

The job postings indicate that DeepSeek is opening two key positions: Harness Product Manager and Harness R&D Engineer, with the work location currently limited to Beijing. DeepSeek's office in Beijing is located in the Haidian District Rongke Information Center, very close to Peking University and Tsinghua University. Officially, it is located in the "Centennial Beijing-Zhang AI Innovation Belt," while informally, it is also in the recently popular "Wang Huiwen Area."

Core Definition: Model + Harness = Agent

In the job description, a core formula is prominently displayed:
Model + Harness = Agent.

This phrase can almost be seen as DeepSeek's internal definition of the productization path for the next stage: the model itself is merely the foundation of the Agent, while context management, tool invocation, task planning, file reading and writing, code modification, terminal execution, feedback collection, and evaluation loops outside the model are the key parts that allow the Agent to truly enter the workflow.

The job posting further states, "We are transforming DeepSeek's cutting-edge model capabilities into leading Agent products. All work outside of the model itself falls under the scope of Harness." Additionally, this position will participate in the entire process of "DeepSeek desktop Agent product" and "define DeepSeek's understanding of Harness."

"Jiazi Light Year" analyzes that DeepSeek is not simply trying to create a code assistant plugin, but is filling in the intermediate layer that connects the model to the real workflow.

Over the past year, the industry has proven that strong coding capabilities do not equate to developers actually using them; a model that can write code does not mean it can continuously complete an engineering task.

What truly changes the way developers work is not a standalone Claude model, but Claude Code; not a standalone GPT model, but Codex; not a code answer in a chat box, but an engineering intelligent agent that can enter the terminal, understand projects, read and write files, run commands, fix errors, manage Git, and invoke tools.

DeepSeek's greatest strength in the past was the model. Now, it is starting to add that layer of "hands" above the model.

I. Why DeepSeek Emphasizes Harness

In the context of traditional AI products, "code assistant" usually refers to two types of products: one is a completion plugin in an IDE, and the other is code Q&A in a chat box.

However, the term that repeatedly appears in DeepSeek's recruitment is not Code Assistant, but Harness.

Harness originally refers to "test harness" or "execution framework" in the engineering context, and in the Agent context, it is closer to a system that truly enables the model to take action. The model is responsible for understanding, reasoning, and generating, while Harness is responsible for integrating these capabilities into the real environment.

The job description mentions that this role needs to plan the DeepSeek Harness product roadmap, connect researchers, engineers, open-source communities, and end-users, and communicate deeply with researchers from the model training team to achieve the co-evolution of the model and Harness.

This statement is crucial.

It indicates that what DeepSeek wants to do is not just to wrap the existing model in a shell, but to make the Agent product itself a part of the model's evolution. In the past, the common product logic for large model companies was: the research team first trains a model, and then the product team creates applications based on the model's capabilities. However, in the Agent era, this order is being disrupted. Products are no longer just an outlet for model capabilities, but a training ground for those capabilities.

A code Agent failing in a real project may not be due to product interaction issues, but rather the model's incorrect compression of long contexts; it may not be a problem with the tool invocation chain, but rather the model's unstable strategy for task decomposition; or it may not be a lack of coding ability, but rather its failure to maintain a continuous understanding of engineering constraints, testing feedback, and user intent.

Therefore, the value of the Harness team is not just to "make products," but to turn real development tasks into a feedback source for the model's continuous evolution.

II. Why DeepSeek Must Fill the Code Harness Gap

DeepSeek has long bet on coding capabilities. From DeepSeek-Coder to DeepSeek-Coder-V2, DeepSeek's investment in code models has continued to increase, supporting improvements in language, context length, and complex task capabilities. Its issue is not the lack of coding ability, but rather that this capability has largely remained at the model level and has not yet become a high-frequency product in developers' daily workflows.

The popularity of Claude Code proves one thing: The competition in AI coding is shifting from competition in model capabilities to competition for entry points into developers' workflows.

This is also a lesson that DeepSeek must learn now. More subtly, before DeepSeek officially took action, the developer community had already created a version of "DeepSeek's Claude Code."

An open-source project named DeepSeek-TUI previously became popular in the developer community. It is a coding agent running in the terminal that can read and write files, execute shell commands, search the web, manage Git, and coordinate sub-agents through a TUI interface.

The rise of DeepSeek-TUI indicates two issues:

1. Basic Mentality Maturity: The DeepSeek model has already established a foundation in developers' minds for becoming a code Agent. Otherwise, the community would not have naturally grown around it to create Claude Code-like products.

2. Official Level Absence: What DeepSeek lacks is not model attention, but official Harness.

In the eyes of developers, the appeal of DeepSeek-TUI is straightforward: low cost, available domestically, long context, and relatively low deployment barriers. Many domestic developers do not want to use Claude Code, but are limited by price, access stability, account systems, and corporate compliance.

However, community projects also have inherent boundaries:

• No matter how active a third-party open-source project is, it is difficult to truly grasp the evolutionary rhythm of the model's internal capabilities;

• It can adapt around the API, but cannot reverse-decide how the model is trained;

• It can optimize prompts, toolchains, and interactions, but it is challenging to systematically inject massive real task feedback into model improvements.

The significance of official Harness lies precisely here.

DeepSeek's own Code Harness has several advantages that community projects do not possess: collaboration with the model team, interface design authority, training data feedback loops, internal real task scenarios, and long-term operational capabilities for the developer ecosystem.

The open-source community has already paved the way: developers indeed need a DeepSeek version of Claude Code. Now, DeepSeek needs to reclaim this path and turn it into its core product.

The fact that DeepSeek is now starting to recruit means it is finally ready to step into the arena itself.

Chen Deli mentioned at the 2025 World Internet Conference in Wuzhen last November: "One of our core advantages is long-termism, insisting on making breakthroughs in cutting-edge intelligence as the main line. In this process, we have also abandoned many side projects and do not engage in those short-term, quick-return side endeavors."

After the model wars, the real Agent wars have begun. What DeepSeek needs to fill this time is the most critical layer between the model and action—Harness.

DeepSeek is equipping its model with a pair of hands.