China’s DeepSeek beefs up agentic AI with ‘harness’ tests as V4 model jolts Silicon Valley

Chinese AI firm DeepSeek is intensifying its foray into agentic AI technology by inviting open-source developers to beta test its forthcoming "harness" software, according to a social media update by Cui Tianyi, who leads the harness team. The software aims to transform large language models, such as DeepSeek's latest V4 Flash model, into autonomous agents capable of executing multi-step tasks and complex workflows.

Agentic frameworks have become a focal point in the AI industry following Anthropic's successful release of Claude Code.

DeepSeek's Cui Tianyi, who joined the company in March to spearhead the harness project, stated that while the team harbored ambitious goals, it was currently understaffed and actively recruiting across multiple platforms. The agentic AI push coincides with DeepSeek's CEO and co-founder, Liang Wenfeng, focusing on cost-effective, high-capacity models as a strategy towards achieving artificial general intelligence (AGI).

This approach has sparked concerns in Washington and Silicon Valley about the growing influence of Chinese open-weight models.

Recently, DeepSeek unveiled "V4 Flash 0731", a 284-billion-parameter lightweight iteration of its V4 series, which garnered praise for its performance comparable to OpenAI's lightweight GPT-5.6 Luna at a significantly lower cost. The new model is priced at US$0.03 per task, making it approximately 105 times cheaper than Anthropic's Claude Fable 5, which costs US$3.15 per task. This pricing advantage has prompted remarks such as "Why pay $1 when I can pay $105" from venture capitalist Chamath Palihapitiya.