DeepSeek Super-agent prototype takes aim at Claude Code

China's best-known AI lab wants its agents to need less babysitting. The DeepSeek Super-agent prototype, released on GitHub after a WeChat post by the company's Harness team on Saturday, is a coding agent built to pick up work from a team's conversations and hand back finished pull requests.

What happened

The repository, dsh-superagent, is published under an MIT licence and describes itself as a research prototype rather than an official DeepSeek product. It works on a single GitHub repository and can optionally listen to Feishu, ByteDance's workplace chat app. When an approved person asks for work, it opens a task, handles it in its own session and directory, and replies in the original thread. A web panel shows what still needs a human.

The design principle is spelled out in the manual: "Take as little human attention as possible, and give as much value as possible." The project acts only when asked or when a rule written by the team allows it, according to its philosophy document as summarised by The Next Web.

The South China Morning Post said the accompanying technical article proposes a way to reduce the need for humans to coordinate several agents and check on their progress. The Straits Times, citing Bloomberg, said the tool can proactively monitor developer group chats and suggest fixes without being summoned.

Why it matters

Coding agents are the most commercially valuable corner of AI right now, and DeepSeek is aiming squarely at Anthropic's Claude Code, Bloomberg reported. DeepSeek created the Harness team earlier this year to focus on that market, the Straits Times said.

Money is not the constraint it once was. The Straits Times said DeepSeek raised about 100 billion yuan in its latest round, one of China's largest, which our DeepSeek funding round report covered. Rivals are also raising at enormous valuations, though our OpenAI funding cut story showed investors growing more selective.

The hard part is coordination. The Next Web pointed to a January CooperBench study in which coding agents working as a team had success rates about 30% lower on average than agents working alone; that study did not test DeepSeek's system. GitHub showed its own multi-session Copilot app at GitHub Universe last week.

The Next Web noted that the public code differs from DeepSeek's internal version and that parts of the design described in the article have yet to be implemented, according to the README.

Developer interest was quick. The DeepSeek Super-agent prototype showed about 9,200 stars and more than 900 forks on GitHub when we checked the repository, figures that measure curiosity rather than real deployment. Its documentation includes an architecture guide, a cookbook for adding new message sources and a troubleshooting section, which suggests the team expects outside developers to experiment with it.

What's next

Watch for adoption on GitHub, independent benchmarks and whether DeepSeek turns the work into a supported product. The project repository includes setup steps and design notes.

The DeepSeek Super-agent prototype also arrives as the company faces limits in the U.S.; The Next Web noted that civilian federal agencies were told to remove DeepSeek last month.

The README calls it research code rather than production-ready software and tells users to review and adapt it first.