Reported by | NUPIAO Staff
On August 12, NUPIAO noticed that the “DeepSeek Harness Team” WeChat official account has recently completed its registration. The account’s logo features a black whale, and it is verified under the entity Beijing DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., with its IP location registered in Haidian, Beijing. As of now, the account has not published any content.
This marks the first time DeepSeek has set up a dedicated official communication channel for its Harness business line. Operating independently from the main account signals that the team wants to build a dedicated platform for announcements, product updates, and technical sharing around this new venture. Industry watchers have interpreted this move as a strong hint that the Harness product is gearing up for an official launch.
Within DeepSeek, Harness is defined as the execution framework for AI agents. It’s essentially a complete control and orchestration system built around large language models — responsible for coordinating tools, managing task states, and closing the execution loop. The goal is to turn a model’s reasoning power into an agent that can actually get complex, real-world things done.
In official job postings, DeepSeek lays out a simple formula: “Model + Harness = Agent.” The company elaborates: “We are transforming DeepSeek’s cutting-edge model capabilities into leading Agent products. Everything beyond the model itself falls under the scope of Harness.”
In plain terms, DeepSeek’s foundational large model acts as the agent’s “brain,” handling reasoning, understanding user needs, and generating code. Harness, meanwhile, is the “executive nervous system” that bridges the gap between that brain and the real environment. A large model can only output text; Harness is what gives AI the ability to actually roll up its sleeves and take action. That’s the core difference between this and your run-of-the-mill chatbot.
Back in May, Chen Deli, a senior researcher at DeepSeek, confirmed that the company was assembling an Agent Harness team internally, with the goal of building DeepSeek Code Harness from the ground up — a direct competitor to Anthropic’s Claude Code. DeepSeek also posted relevant openings on recruiting platforms, all based in Beijing.
The Harness team is led by Cui Tianyi. Public records show that Cui, born in the 1990s, studied computer science at Zhejiang University’s College of Computer Science — making him an alumnus of DeepSeek founder Liang Wenfeng. During his undergrad years, he racked up six gold medals at ACM-ICPC Asia regional competitions. In 2013, he joined the Hong Kong office of Jane Street Capital, one of the world’s top quantitative trading firms, as an assistant quantitative researcher. He stayed there for nine years, working on quantitative systems, risk controls, and backtesting platforms for both equities and fixed income.
In 2022, Cui co-founded TSY Capital, a Hong Kong-based quantitative investment firm, where he led the development of systematic quant strategies. In February 2026, he left TSY, and officially joined DeepSeek in March of this year to spearhead the independent Harness business line — overseeing the overall development of the code agent framework, team building, and product rollout.
This past June, he posted a public recruiting call on overseas social media, saying: “As a newly formed department, the DeepSeek Harness team has ambitious goals and a heavy workload — we’re seriously short-handed. I’ve been interviewing every day and putting up little ads everywhere… We have three open positions: Harness Researcher, Harness Engineer, and Harness Product Manager.”
According to him, while the number of openings had grown, the hiring process was no different from other DeepSeek teams — typically one written test followed by three rounds of interviews, with him personally conducting the final round.
On the recruiting poster for R&D engineers, applicants are expected to have at least two years of software development experience (with exceptions for exceptional candidates), a bachelor’s degree or higher, proficiency in using AI Agent tools for development, and a strong capacity for rapid learning. The role also requires candidates to be heavy users of Agents themselves, familiar with LLM fundamentals — including API, KV Cache, Agent Loop, and so on — and to have solid Chinese communication skills.
The Chinese language requirement was interpreted by some outsiders as “not hiring foreigners.” Cui pushed back on that: “Just as comparable U.S. companies require employees to work in English, DeepSeek requires employees to work in Chinese.” He emphasized, “There is no rule against hiring foreigners.”
In early August, Cui also opened up internal testing recruitment for Harness to global developers, prioritizing applicants with hands-on experience in Agent development — a clear sign that product beta testing is already underway.
Looking at the broader AI landscape, the competitive focus in 2026 has shifted dramatically. With the “war of a hundred models” largely in the rearview mirror, AI agents have become the new battleground. Overseas, Anthropic has rolled out Claude Code, and Google is iterating on Gemini Agent. Domestically, companies like Alibaba, Tencent, and Zhipu have all launched agent products targeting office and coding scenarios. The competition is no longer about leaderboard scores — it’s about the ability to complete complex tasks in real-world environments.
DeepSeek enters the agent race holding a substantial developer base. According to data from OpenRouter, a third-party model aggregation platform, DeepSeek V4 Flash topped the global weekly rankings for AI model API calls between July 27 and August 2, with a staggering 7.22 trillion tokens. A large number of overseas developers are already building secondary applications on top of DeepSeek’s foundational models.
That said, code agents demand extremely high standards for system stability and security boundary controls. Agents have permissions to manipulate files and execute commands — if they overstep, the risks include code tampering and unintended system operations. Building robust safety guardrails is a monumental amount of work.
Beyond that, agent products also need comprehensive documentation, a thriving developer ecosystem, and solid technical support. Building out that entire ecosystem takes time. As one technical expert puts it, having a strong foundational model doesn’t automatically mean you’ll have a strong agent product. The execution framework, prompt engineering, and error rollback mechanisms — a weakness in any single link can cause an Agent to fail in real-world scenarios.
Of course, registering a WeChat account doesn’t mean the product is launching tomorrow. The entire industry is watching to see whether a domestic code agent can actually deliver performance that holds its own against leading overseas products on the global stage.
