DeepSeek unveils AI ‘super agent’ in coding race with OpenAI, Anthropic

Chinese AI firm DeepSeek has unveiled plans for an autonomous AI agent, dubbed the "super agent," as it intensifies competition with US counterparts OpenAI and Anthropic in the development of coding assistants. The super agent aims to streamline complex workflows by reducing the need for human coordination among multiple AI agents and monitoring their progress, according to a technical article by DeepSeek's Harness team.

While AI models have enabled agents to automate diverse tasks, managing multiple agents simultaneously still requires substantial human attention, as a single person can typically oversee only three to five active sessions. DeepSeek's proposed system would automatically allocate tasks among agents and coordinate their work, allowing humans to focus on decision-making and result review.

The initiative aims to elevate agent automation by enabling multiple agents to work simultaneously and respond to events like messages and code updates without explicit human instructions. In one scenario, an agent could independently detect a software bug, attempt to fix it, and submit a proposed code change for human review if it identifies a viable solution.

Experimental tests showed the approach reduced task completion time by nearly 14 minutes compared to traditional methods. While industry experts praised DeepSeek's progress, some questioned whether the proposal offered meaningful advancements over existing systems, emphasizing the importance of ensuring reliability in AI software development.

DeepSeek acknowledged potential security and privacy risks, warning that malicious instructions could compromise connected agents and systems. The company released an open-source research prototype of the super agent on GitHub, positioning it as an architectural exploration rather than a fully developed product.