DeepSeek is putting engineering effort into agent software of the kind Claude Code sells, a product line beyond low-cost model access as it reportedly raises about RMB100 billion.
DeepSeek has published a design for coding agents that keep working in the background and step in only when a person needs to act, and it has open-sourced a prototype. The Hangzhou lab’s Harness team set out the design in a technical article on its WeChat account on the evening of October 31, titled "Exploring Super-agent Design: Towards a More Autonomous Agent System" (Super-agent设计探索：迈向更自主的Agent系统), according to HK01. The code is on GitHub under an MIT license, which lets anyone use and modify it.
Investing.com framed the release as a challenge to Anthropic’s Claude Code, the coding assistant to which programmers delegate development work. DeepSeek’s own materials are more cautious. The prototype’s repository README describes it as a research prototype for reference only and says the published version differs from both the internal version and the paper.
The release still shows where DeepSeek is spending engineering effort while it raises money. Investing.com noted that DeepSeek reportedly raised about RMB100 billion in its latest round, which EastFrontier covered as it neared completion with CATL and Tencent as the biggest backers.
The prototype works on one GitHub repository and acts only on request or by rule
According to the README, dsh-superagent is a proactive coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness, the open-source agent platform the company opened to developers in August, as TechNode reported at the time. It listens to GitHub and, optionally, to the Feishu workplace chat app. When an approved person asks for work, it opens a work item, does the job in its own session and working directory, and answers in the original conversation, for example with a pull request. A web panel lists the items that need a human.
Two operating rules define the design. The agent acts only when an allowed person asks or when a rule written by the team permits it, and it speaks only when someone must act or can use a result. The README sets the goal as taking as little human attention as possible while delivering as much value as possible. It also warns that much of the code was written with coding agents and may contain bugs.
HK01 reported that DeepSeek already runs internal Super-agent instances in its daily development, fixing bugs, responding to code review and iterating on Harness itself.
DeepSeek says human attention has become the bottleneck
"Single-task limits depend on model capabilities, while multitask limits depend on human attention," the paper said, according to the South China Morning Post. DeepSeek argues that a person can typically keep track of only three to five active agent sessions at once, so human attention, not model capability, now limits multitasking. Today’s agents are built around a single session, which means work blocks while waiting, context from different tasks gets mixed, and jobs stop when a device goes offline, according to Zhidx, which summarized the team’s paper.
The proposed answer has two layers. A resident Router takes in messages from sources such as group chats, GitHub and email and dispatches work without doing tasks itself. Isolated Workers carry out tasks in parallel, each with its own context. Environment events act as triggers, filtered first by fast program rules and then, where needed, by a model’s judgment. An external watchdog checks whether Workers are making real progress. The Router can often run on a lighter model, while Workers need the strongest model available, Zhidx reported. In chat apps, replies are capped at 100 words.
The team also published early test results on synthetic workloads of about 300 messages per run, roughly 20 of them work requests, repeated five times. Super-agent finished with a median time of 12.4 minutes and a 95th-percentile time of 15.7 minutes, against a 95th-percentile time of 29.5 minutes for a single session using a subagent tool. Batching cut average model inference time per routing decision from 6.4 seconds to 1.6 seconds. DeepSeek called these results preliminary.
Real teams will show whether the design holds up
DeepSeek’s paper lists four unsolved problems: scaling the two-layer design, working with people outside chat sessions, context boundaries and agent-to-agent communication, and security and privacy. Security matters most for companies weighing adoption, because an agent that watches team chats and repositories sees code and conversations that many employers keep closely held.
Not everyone sees speed as the main problem. "The challenge in AI software development today is ensuring reliability … rather than completing tasks faster," Beijing-based AI entrepreneur Liu Ming wrote on Zhihu, SCMP reported.
The prototype now goes to developers who can run it on their own repositories and report whether the attention savings appear outside DeepSeek’s test setup. The signal to watch is whether DeepSeek folds Super-agent into a supported Harness release with paid model usage behind it, which would turn a research prototype into a product line competing directly with Claude Code.