DeepSeek prepares new AI agent model to rival US competitors
By designing AI that adapts from past actions, DeepSeek hopes to match the industry’s goal of reducing human supervision in complex digital tasks.
According to people familiar with the plans, Chinese startup DeepSeek is developing an AI model with enhanced agent features to compete with US firms such as OpenAI.
The Hangzhou-based company intends for the system to perform multi-step tasks with limited input and adapt from its previous actions.
Founder Liang Wenfeng has urged his team to prepare the release before the end of 2025. The project follows DeepSeek’s earlier success with R1, a reasoning-focused model launched in January that attracted attention for its low development costs.
Since then, DeepSeek has delivered only incremental updates while rivals in China and the US have accelerated new product launches.
The shift towards AI agents reflects a broader industry move to develop tools capable of managing complex real-world tasks, from research to coding, with less reliance on users. OpenAI, Anthropic, Microsoft, and Manus AI have already introduced similar projects.
Most systems still require significant oversight, highlighting the challenges of building fully autonomous agents.
DeepSeek declined to comment on the development.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
