The DeepSeek intelligent agent products are coming.

DeepSeek is advancing the productization of Agents.

Recently, DeepSeek has successively posted two positions related to Agent Harness: Agent Harness Product Manager and Agent Harness R & D Engineer.

According to the job description, DeepSeek is transforming cutting - edge model capabilities into leading Agent products. All work except the model itself falls within the scope of Harness.

New hires will join the Harness team, collaborate with engineers, researchers, and product managers, participate in the entire process of developing DeepSeek's desktop Agent product, and define DeepSeek's understanding of Harness.

Among them, the requirements for Agent Harness R & D engineers are to be high - intensity users of Agent products, have great enthusiasm for the development of Agent Harness, have good taste and judgment in model behavior, and have a strong sense of developer experience. They should have used code - based and general - purpose Agent products extensively and integrated the use of relevant products into their work and life.

Agent Harness Product Managers should be familiar with various forms of Agent products, have used products similar to Claude Code, Cowork, Codex, Cursor, OpenCode, GitHub Copilot, Manus, Openclaw, Hermes, etc. extensively, and also integrate the use of relevant products into their work and life.

Previously, media reported that DeepSeek is organizing a new Harness team internally, focusing on programming agent products, with an internal benchmark of Claude Code under Anthropic.

Chen Deli, a senior researcher at DeepSeek, also confirmed this in a post on social media. He said that "DeepSeek is organizing a new Harness team to work on products and research in the Harness direction" and directly stated that "simply put, it's about benchmarking ClaudeCode and creating DeepSeek CodeHarness."

The Harness architecture is an engineering infrastructure reference framework in the field of agents. It aims to transform the raw intelligence of foundation large models into reliable, controllable, and usable agent capabilities through a systematic engineering solution, which can systematically make up for the inherent deficiencies of naked models in memory, code execution, tool invocation, etc. This concept was promoted by Anthropic from late 2025 to early 2026 and has received wide attention, marking an important evolution of AI engineering practice from prompt engineering and context engineering to building a complete agent operating environment.

DeepSeek has always been betting on AI programming capabilities. On April 24th, DeepSeek released a preview version of its latest large - language model V4, stating that it excels in agent tasks, knowledge processing, and reasoning capabilities, and has carried out special optimization for mainstream agent tools such as Claude Code under Anthropic.

After that, DeepSeek launched the image recognition mode in a limited - release manner, which addresses the problem that after DeepSeek V4 - Pro is connected to Claude Code, it has difficulty recognizing architecture diagrams, design drafts, and error screenshots during daily development.

Currently, the global daily average token consumption has exceeded 360 trillion. According to IDC's prediction, it will increase by 300 million times in the next five years. Among them, AI programming and agents are one of the main consumers. Alibaba CEO Wu Yongming revealed at the latest earnings conference that from November last year to May this year, a large part of the growth in Alibaba Cloud's API usage was brought about by AI coding. "Agents can solve complex tasks in digital work, and there will be a growth trend in the next two or three years."

The establishment of DeepSeek's Harness team this time is not only a benchmark in the technical route but also a strategic move in the productization pace. With its model adaptation ability and the advantage of open - source community collaboration, it is worth continuously observing whether DeepSeek can catch up and take the lead.

This article is from the WeChat official account "Science and Technology Innovation Daily", author: Huang Xinyi. Republished by 36Kr with permission.