이미지: METAL LAB 생성
Summary
- DeepSeek has opened an official "DeepSeek Harness Team" account on WeChat and posted job listings
- The team's stated goal is to build AI agents that can compete with Anthropic's Claude Code
- The account is operated by a Beijing-based entity and confirmed to be verified by Tencent
- 신설 조직
- DeepSeek Harness Team
- 채용 플랫폼
- 위챗(WeChat) 공식 계정
- 목표 서비스
- 앤트로픽 클로드 코드 유사 AI 코딩 에이전트
- 운영 법인 소재지
- 베이징(딥시크 본사는 항저우)
- 확인 경로
- 중국 기업 정보 사이트 치차차(Qichacha), 텐센트 인증
A job posting appears on WeChat
DeepSeek has opened a new official account on WeChat and posted job listings under it. The account is named "DeepSeek Harness Team." According to Chinese corporate registry site Qichacha, the account is operated by a Beijing-based entity that is confirmed to be controlled by DeepSeek. The account itself has been verified by Tencent. The stated goal in the posting is clear — to build AI agents that can perform tasks on a user's behalf, much like Anthropic's Claude Code.
What exactly does a "harness" team build?
In the AI industry, a "harness" refers to the tooling layer that wraps around a model. A language model on its own cannot read files, execute code, or operate a terminal. A harness is the framework that attaches capabilities like file access, code execution, and web search to a model so it can carry a task through to completion. Many observers credit Claude Code's popularity not to the underlying model itself but to the strength of its harness design. In other words, what DeepSeek is targeting this time isn't a competition over model performance, but the "last mile" of connecting a model to real-world work.
DeepSeek, headquartered in Hangzhou, has drawn attention for a string of low-cost, high-performance open-source models. On August 9, the company released DeepSeek-V4-Pro — targeting a million-token context window — on Hugging Face under an MIT license. Earlier this month, a benchmark from Together AI comparing V4 Flash and GPT-5.6 Luna found that a DeepSeek-first cascading approach cut per-task costs by 37%. The formation of this new harness team makes it clearer that DeepSeek's ambitions extend beyond model performance into actual task-execution capability.
DeepSeek releases DeepSeek-V4-Pro on Hugging Face
So what does this change?
Claude Code is an agentic tool that, given natural-language instructions from a developer, can write code, test it, and even deploy it on the developer's behalf. If DeepSeek targets this market with an open-source approach, developers could gain room to piece together cheaper alternatives instead of relying on costly commercial agents. Team formation is still at an early stage, and no concrete product or launch timeline has emerged yet, but the move is being read as a signal that DeepSeek intends to expand its footprint from model releases into the competition over agentic tools.
