DeepSeek Harness Team Releases Super-agent Design Exploration and Open-Sources Prototype

By Caijing Circle

On the evening of October 31, the DeepSeek Harness team published a technical article on its official WeChat account titled "Super-agent Design Exploration: Toward a More Autonomous Agent System," and the accompanying prototype dsh-superagent-prototype was open-sourced on GitHub the same day under the MIT license.

The article opens with a self-imposed caveat: this is an architecture-ideas discussion oriented toward engineering practice, not an introduction to existing product features.

But the judgments offered in the article are quite specific: as model capabilities improve, AI increasingly needs no human intervention to complete a single task, yet a person's attention for managing multiple tasks has not grown in step—one person can still only keep up with 3 to 5 active sessions at once. DeepSeek's exact words are: the ceiling for a single task depends on model capability, while the ceiling for multiple tasks depends on human attention.

This sentence is the starting point of the entire article. The article also notes that improvements in model capability have given rise to a batch of new forms, with personal assistants beginning to emerge in concentrated numbers; they connect directly to messaging, calendars, email, and work platforms, continuously observing information flows. This always-online, event-driven way of working is not limited to personal scenarios—it holds equally in commercial scenarios such as software development—but the designs of existing mainstream agents cannot sustain this scale.

Every design decision in the new architecture is aimed at solving this bottleneck.

01 If sessions cannot communicate with each other, the ceiling falls on people

It is understood that existing agents take a single session as the basic unit, and sessions are independent of one another. This leads to a phenomenon: conclusions reached in session A are unknown to the agent in session B. When a user pushes multiple workstreams forward in parallel, they must switch manually and remember every piece of context themselves. In the end, only the human knows the full picture of all tasks.

DeepSeek's solution is to divide labor among AIs. One layer is called the router, a continuously running global coordination node that uniformly receives messages from multiple channels such as group chats, GitHub, and email, maintains shared context, and dispatches tasks without delving into execution itself. The other layer is called the worker; each maintains its own independent context and works in parallel on problem reproduction, code review, and feature development, and can call a Subagent when necessary.

This design differs from the common "main agent with a few sub-agents." The main agent needs to understand task details, plan subtasks, and coordinate at high frequency, which places a heavy reasoning burden on it; the router only handles who a given matter should be pushed to, workers' contexts are isolated from one another, and conflicts are discovered on their own by subscribing to the same topic. After taking over a task, a worker can also subscribe directly to follow-up messages, saving the coordination overhead of forwarding through the router.

DeepSeek itself says this design proposes no entirely new theory; it moves classic engineering ideas from computer systems onto agents.

Event-driven design makes group chat messages, emails, calendar reminders, GitHub comments, and CI test results all events that trigger action. Storage-compute separation puts the execution loop in the cloud and stores session state in a database, leaving only a lightweight daemon locally, so tasks can keep running even when the computer is offline. An external watchdog checks whether a task has fallen into an invalid loop and attempts to repair it. The asynchronous ReActLoop (reasoning-action loop) lets the AI still respond to new instructions while executing time-consuming tasks; this feature is not yet enabled by default.

In addition, there is a layer of speculative execution: when it is still uncertain whether a result will be adopted, the system tries it in the background in advance, submitting it if valuable and abandoning it if not.

Moreover, "proactiveness" is a keyword of this system, and also the place where the most restraint is needed. The vast majority of group chat messages do not need AI involvement, so DeepSeek set up two levels of filtering: first, programmatic rules screen out obviously irrelevant messages, then a large model judges whether the remainder is casual grumbling or a real problem; only after passing both is a task dispatched.

The article gives an internal example: a team member says in a group chat, "this project has a BUG," without anyone @-mentioning it; the Super-agent silently reproduces it in the background, and only after confirming the problem exists and finding a simple fix does it submit a PR (pull request) for a human to review. Later, when someone asks in a GitHub comment for a different fix, it can switch from the chat software's context to GitHub and continue working. This strategy is called "silent reproduction and fix" internally.

The article also provides a set of experimental data: the Super-agent based on this approach had a median task completion time of 12.4 minutes and a 95th percentile of 15.7 minutes, better than an ordinary session with sub-tools. DeepSeek also states that this is a preliminary experiment based on synthetic load and a simplified environment, not a formal product conclusion.

The product boundaries are also stated plainly. The prototype on GitHub is a research prototype, not a formal product. According to the repository description, the publicly released version is not the same as the system DeepSeek runs internally, and some features in the article are still under development. At present it mainly targets a single GitHub repository, and can optionally connect to Feishu. The Super-agent acts only when requested or when rules written in advance by the team permit it; when there is no result it does not narrate the process, and without permission it does not act on its own.

Worth noting is the team behind this. The publisher is the Harness team, not the model team.

02 After two rounds of financing, what it outputs to the outside has changed

Laying out the timeline makes the position of this news clear.

In April 2026, DeepSeek launched the first round of external financing since its founding, completing settlement by the end of June, with a scale of about 51 billion yuan and a post-investment valuation approaching 400 billion yuan. Among the contributors, Liang Wenfeng personally invested about 20 billion yuan, Tencent about 10 billion yuan, the CATL system about 5 billion yuan, and NetEase, JD.com, Lisi Capital, and IDG Capital each about 3 billion yuan; the National Artificial Intelligence Industry Investment Fund was also among them.

Starting October 22, news of a new round of financing spread through the market: a planned raise of about 100 billion yuan, with an initial fundraising target of 50 billion yuan that was raised at least twice thereafter. Tencent and CATL were the largest investors in this round. Multiple media outlets cross-verified from different sources; as of publication, DeepSeek and the relevant investors had made no public official response, and the final amount remained undetermined.

Adding the two rounds together, the cumulative financing scale within half a year was about 150 billion yuan.

The money has destinations. According to public disclosures, in July it participated in the IPO subscription of ChangXin Technology, in August it spent 141 million yuan on a strategic placement in Unitree Robotics, and in October it planned to lead an investment in a robotics startup; earlier reports showed that it was also advancing the deployment in Ulanqab, Inner Mongolia, of a data center with at least 160,000 Huawei Ascend accelerator chips.

An IPO is also being advanced in parallel. Since late August, DeepSeek has approached multiple investment banks; rumors say CITIC Securities has already begun due diligence, and market talk is that it could submit a listing application as early as late 2026 and list in 2027; in late September, Yan Wentao, a former partner at Hillhouse Venture Capital, joined as CFO. A September report by 21 Finance showed that the tutoring agreement had not yet been signed, and that the 2027 timetable was market rumor. None of these claims has been officially confirmed.

Putting these together with Harness points to the same thing: DeepSeek's external label has always been models, and open-source models such as V3 and R1 defined its identity. After financing, in addition to continuing to train models and buy computing power, it began to output "how to get work done"; from previously delivering only a more advanced model, it shifted to delivering a system that schedules, executes, and accepts on people's behalf.

03 The entrance is already crowded, and DeepSeek chose open source

It is worth noting that the Super-agent was chosen for release at the end of October; in the same window, this track was already full of players.

Sina Finance gave a summary in a recent report: competition among agents is moving from executing specific tasks toward continuously providing personal services and becoming true personal agents.

The list of those doing the same thing at the same time is not short. ByteDance released the productivity-oriented agent brand "Doubao Work" on August 25, deeply integrated with Feishu, allowing users, after authorization, to operate computers across software; Tencent's Workbuddy had long been in trial operation; Alibaba Qwen announced accelerated efforts to build a Personal Agent; Ant's Lingguang opened beta testing for a new version.

Moonshot took another path: while iterating Kimi to 3 trillion parameters and launching a Pro version for the first time, it also launched the "Kimi Enterprise Partner Program" in September, bringing in IT service providers such as Teamsun and Kingsoft Cloud to help enterprises deploy AI into core business.

Actions on the enterprise side are also intensifying. Tencent said on October 29 that its agent development platform ADP allows enterprises to customize, manage, and coordinate agents for customer service, marketing, and operations; Alibaba Cloud launched Agent Loop and Agent Teams in July, together with the existing Agent Run, for agent development, deployment, and monitoring.

English text is being prepared.

This posture has a direct impact on the application layer. In the past, the moat of the application layer was "knowing better how to invoke models"; once underlying model companies open-source agent frameworks and best practices, this capability gap will rapidly thin. For DeepSeek itself, it is also a new battlefield: previously it compared models only with OpenAI and Anthropic; now it must compare products, scenarios, and who can truly put agents into real workflows with ByteDance, Tencent, and Alibaba.

One sentence in the article is DeepSeek's positioning of itself: "The system takes over tedious session management and context synchronization at the underlying level, enabling humans to move further up from specific execution details to higher-level decision-making and acceptance."

This is the company's definition of its own product, not a judgment on the industry. To test it, two things must be watched in the future: how long that internal instance has been running and what it has achieved; and whether the publicly released prototype can expand from "a GitHub repository" to more scenarios.

Finally, the article does not disclose the operating scale of that internal instance, nor does it give a task success rate. The public part is a research prototype, a GitHub repository, and a set of experimental data based on synthetic load. Therefore, everything remains to be tested.