Farsight News, August 11 - Beijing DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. has taken a step forward in its AI endeavors by registering a WeChat official account named 'DeepSeek Harness Team.' As of now, no articles have been published on this newly established platform.

In the realm of AI, the term 'Harness' denotes an intermediate scheduling and adaptation layer that operates between the foundational large model and the upper-level agent. DeepSeek's creation of this dedicated account suggests its ambition to delve into the realm of code agents.

Interestingly, three months prior, 'Jiazi Guangnian' reported that DeepSeek had internally assembled a new Harness team, with a specific focus on developing code agent products that would rival Anthropic's Claude Code. This development was further corroborated by concurrent job postings, which revealed that DeepSeek was actively recruiting for two key positions: Harness Product Manager and Harness Research and Development Engineer, both to be based in Beijing.

DeepSeek's decision to venture into the development of its own code agents is primarily motivated by the challenges associated with third-party scheduling frameworks, such as context decay, tool call confusion, and cache failure. By developing its in-house Harness, DeepSeek aims to customize the solution to its model's unique characteristics, thereby optimizing KV caching, long-text compression, and session memory. This, in turn, is expected to substantially reduce inference costs and enhance the success rate for complex tasks.

Li Nan, the former Vice President of Meizu Technology, expressed optimism regarding DeepSeek's move to develop its own Harness. He stated, 'If DeepSeek V4 Pro is integrated with DeepSeek Harness, its performance could closely match, or even surpass, that of Kimi K3 + Kimi Work, all while being more cost-effective.'