TLDR

  • DeepSeek published a design for AI coding agents that need less human input.
  • The company also released a prototype called Super-Agent, reported by Bloomberg.
  • A person can only oversee three to five AI agent sessions at once, the paper said.
  • DeepSeek reportedly raised about $15bn in one of China’s largest private funding rounds.
  • IPO filing could come by the end of 2026 or early 2027, according to anonymous sources.

Chinese AI company DeepSeek has revealed a design to make AI coding assistants more independent. The goal is to let them handle several tasks with less human help.

The plan comes from a paper by the company’s DeepSeek Harness team. The South China Morning Post reported on it on Sunday.

Why Human Attention Is the Limit

The team said managing many agents takes a lot of human attention. It wrote that “single-task limits depend on model capabilities, while multitask limits depend on human attention.”

The paper said one person can oversee only three to five active agent sessions at the same time. That is the problem the design tries to solve.

Under the proposed system, tasks would be shared among agents, and the agents would coordinate their own work. People would then be free to review results and make decisions.

The SCMP said the proposal comes as competition grows among AI developers. These include US firms Anthropic and OpenAI.

Super-Agent Prototype Released

DeepSeek has also released a prototype agent called Super-Agent. Bloomberg reported the details, citing a post by the DeepSeek Harness team on WeChat.

The team said the open-source agent can watch group chats among developers. It looks for feedback on coding issues and suggests fixes without being asked.

The Harness team was created earlier this year. It was reportedly set up to build AI agents that can compete with those from rival firms.

Cointelegraph reported last October that Chinese AI models were beating US models in crypto trading. The data came from analytics platform CoinGlass.

In that experiment, DeepSeek was the only AI model with a positive unrealized return, at 9.1%. The result surprised traders because DeepSeek was built at a fraction of the cost of US rivals.

Anthropic filed for an IPO in June this year. OpenAI did the same one week later.

OpenAI later halted its plans over AI safety concerns. CEO Sam Altman said 2026 was no longer the right time to go public.

Anthropic is nearing its IPO. It is expected to raise $100bn or more, which would be one of the largest listings in history.

In late October, reports said DeepSeek raised around $15bn in one of China’s largest private fundraising rounds. Investors included Tencent and battery maker Contemporary Amperex Technology.

The raise reportedly came ahead of IPO preparations. Anonymous sources told publications DeepSeek could file by the end of the year or in early 2027.