DeepSeek is making a quiet but direct move into AI coding agents. While the market has been waiting to see when DeepSeek Harness will formally appear, the company has already opened a WeChat account named "DeepSeek Harness Team" and used recruiting posts to signal what it is building: a tool meant to compete with Claude Code and Codex.
A WeChat account points to the project
The account, "DeepSeek Harness Team," is registered in Beijing. According to the report, checks on the Chinese corporate database Qichacha show that it is controlled by DeepSeek, and the account has also received Tencent’s official verification.
There was no launch event and no formal standalone announcement. Instead, DeepSeek let the message come through job postings. One recruitment notice described the goal in plain terms: turn the company’s models into a leading-edge agent product and compete with Anthropic and OpenAI for the same group of enterprise clients.
The effort was already underway
The report says this did not start recently. DeepSeek senior researcher Chen Deli had previously confirmed on social media: "Benchmark Claude Code and build DeepSeek Code Harness." The team is being led by Cui Tianyi.
In comments made public at the end of June, Cui said the unit had only just been formed, its ambitions were large, and it was severely understaffed. He said he was interviewing candidates every day and posting recruitment messages across platforms. On Aug. 1, Cui went a step further and publicly called for global beta testers, giving priority to developers with experience in open-source Agent Harness projects. Applicants were asked to submit their code hosting platform accounts and representative work.
What a harness does
The report describes a harness as the software framework wrapped around a large language model. Its job is to convert the model’s raw reasoning ability into an agent that can break down tasks, call tools, and carry out multi-step work in sequence.
Put another way, if the model is the brain, the harness provides the hands, feet, and operating process. It determines how the system reads files, runs tests, and corrects itself after mistakes instead of simply responding one prompt at a time.
That is why Claude Code became a product story for Anthropic, the report argues. The differentiator was not only the intelligence of the Claude model itself, but also the harness that translated that intelligence into the ability to complete work. In practice, the industry view cited in the report is that the same model can deliver very different user experiences depending on the framework built around it, and that this layer is often what decides whether users stay.
Competition shifts from model scores to product experience
By making its Harness team public now, DeepSeek is showing where it wants to compete next. The report frames that move as a sign that Chinese AI players are no longer focused only on matching model benchmark scores, but are pushing into product experience as well.
The same report also says DeepSeek upgraded the agent capabilities of its flagship V4 Pro model in the same week, taking aim at a business that could reach $8 billion in annual revenue.