Chinese artificial intelligence startup DeepSeek has unveiled a new architecture aimed at making AI agents more autonomous, marking the latest push by China's model developers to move beyond chatbots toward systems capable of independently carrying out complex tasks.

DeepSeek's Harness team said in an article published on Saturday that its experimental "super-agent" architecture is designed to allow AI to continuously monitor information, manage multiple tasks in parallel and take action with less human supervision.

The development points to a broader shift in China's fast-moving AI industry, where competition is increasingly moving beyond model size and benchmark scores toward agents that can plan, execute and complete real-world work.

As AI models become better at handling individual tasks, Deep-Seek said the bottleneck is increasingly human attention: users still have to switch between different conversations, keep track of their progress and provide instructions.

The super-agent approach is intended to move much of that coordination into the AI system itself, allowing people to focus more on higher-level decisions and reviewing results rather than overseeing each step of a task. Deep-Seek stressed that the architecture remains an engineering exploration rather than a feature available in its existing products.

DeepSeek said an internal version of the system is already operating continuously. When employees flag a software problem in a group chat, for example, the AI can identify the potential defect even without being directly addressed, attempt to reproduce the problem in the background and search for a fix.

The push reflects a growing focus across the AI industry on so-called agentic AI — systems designed not only to generate text or answer questions, but also to make plans, use tools and carry out multi-step tasks.

Chinese developers have accelerated model releases in recent weeks. The flurry of releases suggests China's AI race is entering a new phase. Rather than simply building models that are bigger or score higher on tests, developers are increasingly competing over whether AI can reliably turn intelligence into action.

Moonshot AI released last month Kimi K3.1 Pro, the first model in its new K3.1 series. The company said the model, its most capable in the roughly 3-trillion-parameter class, offers substantial improvements in coding, 3D creation, financial research and other long-horizon agent tasks, while supporting a context window of 1 million tokens.

MiniMax launched its M3.1 Flash model mid-October, saying it had strengthened performance in agentic tasks, coding and computer use while delivering online inference speeds nearly three times faster than its predecessor and cutting the overall cost of agent tasks by 70 percent.

DeepSeek itself updated its V4.1 Flash model on Oct 17. DeepSeek said the update improved performance while allowing tasks to be completed faster and with shorter outputs.

Wei Kai, head of the AI research institute at the China Academy of Information and Communications Technology, said in a separate interview that agents represent an engineering enhancement built on top of large language models that can significantly extend their capabilities.

But challenges remain. "Agents still need improvement in reliability, contextual memory and long-horizon tasks, and there is still some distance to go before large-scale adoption," Wei said.

chengyu@chinadaily.com.cn