Author | Wang Bo, Jiazi Guangnian
"Jiazi Guangnian" learned from insiders close to DeepSeek that DeepSeek is internally forming a new Harness team focused on code agents, internally benchmarked against Anthropic’s Claude Code.
Recently, DeepSeek senior researcher Chen Deli also confirmed this on social media, stating that "DeepSeek is forming a new Harness team to develop products and conduct research in the Harness domain," and bluntly added, "In simple terms, we're building DeepSeek Code Harness to compete with Claude Code."
This is not a regular hiring process.
The job posting reveals that DeepSeek is currently opening two key positions: Product Manager for Harness and Engineering Manager for Harness, with the work location currently limited to Beijing. DeepSeek’s Beijing office is located in the Rongke Information Center in Haidian District, close to both Peking University and Tsinghua University. Officially, this area is referred to as the "Century Jingzhang AI Innovation Corridor," while colloquially, it is also known as the recently popular "Wang Huiwen Area."
Core definition: Model + Harness = Agent
In the job description, a core formula is placed in the most prominent position:
This statement can almost be regarded as DeepSeek’s internal definition of its next-stage productization path: the model itself is merely the foundation of the agent; it is the components beyond the model—context management, tool invocation, task planning, file reading and writing, code modification, terminal execution, feedback collection, and evaluation闭环—that are crucial for the agent to truly integrate into workflows.
The job posting further states: “We are transforming DeepSeek’s cutting-edge model capabilities into leading Agent products. All work beyond the models themselves falls within Harness’s scope.” Additionally, this role will participate in the entire process of developing the DeepSeek desktop Agent product and will “define DeepSeek’s understanding of Harness.”
Jiazi Guangnian analyzes that DeepSeek is not simply aiming to create a code assistant plugin, but rather is filling in the intermediate layer needed to connect models to real-world workflows.
Over the past year, the industry has demonstrated that strong coding ability does not mean developers will actually use it; and a model’s ability to write code does not mean it can consistently complete an engineering task.
What truly transforms how developers work is not the standalone Claude model, but Claude Code; not the standalone GPT model, but Codex; not code answers in a chat box, but an engineering agent that can access the terminal, understand projects, read and write files, run commands, fix errors, manage Git, and invoke tools.
DeepSeek was previously strongest in its models. Now, it is beginning to add the “hand” on top of the models.
I. Why does DeepSeek emphasize Harness?
In the context of traditional AI products, "code assistant" typically refers to two types of products: completion plugins within IDEs and code Q&A chatbots.
But the term that repeatedly appeared in DeepSeek's recent hiring efforts was not "Code Assistant," but "Harness."
Originally, "Harness" in engineering contexts refers to a "test harness" or "execution framework"; in the context of agents, it more closely resembles an external system that enables the model to take action. The model handles understanding, reasoning, and generation, while the Harness connects these capabilities to real-world environments.
The job description states that this role requires planning the product roadmap for DeepSeek Harness, connecting researchers, engineers, the open-source community, and end users, and engaging in close collaboration with researchers on the model training team to enable the co-evolution of the model and Harness.
This statement is crucial.
It shows that DeepSeek aims to do more than simply wrap existing models in a new interface—instead, it seeks to make the Agent product itself an integral part of the model’s evolution. In the past, the common product logic among large model companies was for research teams to first train a model, followed by product teams building applications based on the model’s capabilities. But in the Agent era, this order is being overturned: products are no longer merely outlets for model capabilities—they are training grounds for those capabilities.
A code agent failing in a real-world project may not be due to product interaction issues, but rather because the model's approach to compressing long contexts is flawed; it may not be a problem with tool invocation pipelines, but rather because the model's strategy for task decomposition is unstable; it may not be a lack of coding ability, but rather a failure to continuously understand engineering constraints, test feedback, and user intent.
Therefore, the value of the Harness team lies not just in “building products,” but in turning real development tasks into a feedback source for continuous model evolution.
II. Why DeepSeek Must Implement Code Harness
DeepSeek has long bet on coding capabilities. From DeepSeek-Coder to DeepSeek-Coder-V2, DeepSeek has consistently increased its investment in code models, continuously enhancing support for languages, context length, and complex task performance. The issue isn’t whether it has coding capabilities, but rather that these capabilities have largely remained at the model level and haven’t yet become high-frequency tools in developers’ daily workflows.
The popularity of Claude Code proves one thing: competition in AI coding is shifting from model capability to competition over entry points into developer workflows.
This is also a lesson DeepSeek must now address. More subtly, before DeepSeek’s official team acted, the developer community had already created their own version of “DeepSeek Claude Code.”
An open-source project called DeepSeek-TUI previously gained popularity among developers. It is a coding agent that runs in the terminal, capable of reading and writing files, executing Shell commands, searching the web, managing Git, and coordinating sub-agents through a TUI interface.
The popularity of DeepSeek-TUI highlights two issues:
Basic mental maturity: The DeepSeek model has already developed the foundational capability to function as a code agent in developers' minds; otherwise, the community would not have naturally evolved products like Claude Code around it.
Lack at the official level: DeepSeek is not lacking model attention, but rather an official harness.
To developers, DeepSeek-TUI’s appeal is straightforward: low cost, availability in China, long context length, and relatively low deployment barriers. Many Chinese developers aren’t unwilling to use Claude Code—they’re constrained by pricing, access stability, account systems, and enterprise compliance.
But community projects also have natural boundaries:
Even the most active third-party open-source project struggles to truly keep pace with the evolution of a model's internal capabilities;
It can be adapted around the API, but it cannot reverse-determine how the model is trained;
It can optimize prompts, toolchains, and interactions, but it is difficult to systematically incorporate vast amounts of real-world task feedback into model improvements.
The significance of the official Harness lies precisely here.
DeepSeek builds its own Code Harness and possesses advantages that community projects lack: collaborative model development, control over interface design, a closed loop for training data, access to real internal use cases, and long-term capabilities for nurturing the developer ecosystem.
The open-source community has already paved the way: developers clearly need a DeepSeek version of Claude Code. Now, DeepSeek is taking back this path and turning it into its own flagship product.
Meanwhile, DeepSeek's official recruitment efforts signal that it is finally ready to enter the arena directly.
In November last year, Chen Deli mentioned at the World Internet Conference Wuzhen Summit: “One of our company’s core strengths is long-termism—we remain committed to the core mission of achieving cutting-edge intelligent breakthroughs. Along the way, we have also let go of many side initiatives, avoiding short-term, quick-profit pursuits.”
After the model war, the real agent war has begun. This time, DeepSeek aims to fill the most critical layer between models and actions—Harness.
DeepSeek is giving its model a pair of hands.