DeepSeek personally went on-site to conduct a competitive analysis of Claude Code, with internal confirmation that the Harness team is currently being formed

According to DynoMatic Beating monitoring, DeepSeek's senior researcher, Chen Deli, confirmed on social media that the company is internally forming a new Harness team, directly benchmarking Anthropic's Claude Code, with a focus on developing the "DeepSeek Code Harness."

The official recruitment page indicates that DeepSeek has currently opened two key positions for a Harness Product Manager and R&D Engineer in Beijing's Haidian District, clearly stating that new team members will be involved in the entire R&D process of the "DeepSeek Desktop Agent Product."

The recruitment description defines the core path as "Model + Harness = Agent." DeepSeek categorizes all engineering actions outside of the model, such as context management, tool invocation, file I/O, terminal execution, and test feedback, under the Harness category.

Prior to this, the developer community had spontaneously launched the open-source terminal project DeepSeek-TUI as a benchmark for the Claude Code. The official formation of the Harness team signifies that DeepSeek is reclaiming control of the terminal entry point, attempting to directly inject code test feedback from real projects into the evolutionary loop of the underlying model.