DeepSeek open-sources Super-agent, a background coding agent prototype

DeepSeek’s Harness team has released dsh-superagent, an open-source prototype of a more autonomous coding agent, publishing the code on GitHub under the MIT licence on October 31. The team says the agent can monitor developer group chats for coding issues and suggest fixes without being summoned, and coverage presents it as a challenge to Anthropic’s Claude Code.

The release pairs the code with a design article, which the team posted on Saturday, arguing that human attention, not model capability, is now the main bottleneck when several agent tasks run at once. DeepSeek is clear that this is not a product: the repository calls itself a “research prototype for reference only”.

What dsh-superagent is and where it comes from

DeepSeek created the Harness team earlier in 2026 to build AI agents that can compete with the coding services of leading US labs, including Anthropic. “Harness” here means the software scaffolding wrapped around a large language model that lets an agent manage more complex tasks.

The team’s main project is DeepSeek Harness (DSH), an open-source program the company uses to bring its models to users’ computers as agents, positioned as an alternative to Claude Code. DeepSeek released a developer preview of DeepSeek Harness in August, describing it as a framework that lets AI models use tools and execute tasks through a modular, plugin-based architecture. The new Super-agent prototype is a plugin bundle for that program.

The accompanying paper is titled “Exploring Super-agent Design: Towards a More Autonomous Agent System” in one English rendering, and “Super-agent design exploration: Towards a more autonomous agent system” in another. DeepSeek announced the work in a WeChat post on October 31, according to The Straits Times, and in a Weibo post cited by Stocktwits.

How the Router and Worker design differs from session-based agents

DeepSeek’s argument starts with how most coding agents work today. A developer gives an instruction, the agent works on it, then waits for the next command, all inside one session. When many tasks run at once, the team says, the human becomes the limit: according to the team, one person can keep track of only three to five active sessions at a time.

Super-agent replaces the single session with two layers. A Router handles communication with users, keeps the shared context and assigns tasks; several Workers carry out assignments independently and report progress and results. DeepSeek says this lets the same agent follow a task as it moves between channels, and that an internal version can track an issue from an instant-messaging conversation into GitHub.

The team lists four goals for the design: continuity, efficiency, proactivity and persistence. In practice that means one agent reachable across messaging and coding platforms, parallel handling of complex tasks, spotting things that need attention without explicit instructions, and staying up for long periods, including recovering from interruptions. The system relies on event-driven operation, meaning it reacts to outside events such as chat messages rather than only direct commands, plus asynchronous interaction and an external monitor called Watchdog.

The stated aim is to let humans “move up from the details of execution to higher-level decisions and acceptance,” in the paper’s framing as quoted by Trending on Weibo.

How the published prototype works, step by step

A review of the repository by Pasquale Pillitteri, who did not run the code, describes the flow of the public version:

1. Source. A stream of events. The GitHub source polls the repository every minute by default and collects notifications for the agent’s account; the Feishu source reads configured chats. Feishu is ByteDance’s enterprise messaging platform, called Lark outside China.
2. Trigger and Thread. Every event becomes an immutable Trigger tied to a conversation, called a Thread.
3. Eval. A first check looks at who is writing and whether they are addressing the agent by mention or direct message. A second has the agent’s main session decide whether to accept the message, ignore it or attach it to work already open.
4. Ticket and Mandate. An accepted message becomes a Ticket, a unit of authorized work, with a Mandate recording why the work is allowed and how far it can go. A Mandate never widens; a bigger scope needs a new Ticket.
5. Worker. Each Ticket gets its own executor session, named after the ticket, with its own working directory. The result, such as a pull request, goes back to the conversation it came from.

Some engineering choices are concrete. Shell commands move to the background after 5 seconds so a Worker can read new messages while tests run; background jobs get at most 30 minutes after each turn before they are terminated and the Ticket reports that it is waiting; and all state lives in a single SQLite file.

Item	Published prototype
Licence	MIT
Language	TypeScript; 111 files under src, 65 under tests
Documentation	Nine numbered documents (000 to 008) and about ten Markdown prompts the agent reads
Event sources	GitHub (polls every minute by default), Feishu
Background command switch	After 5 seconds
Background job limit	30 minutes per turn
State storage	Single SQLite file
Host program	DeepSeek Harness 0.2.1-alpha.2 or later

Internal agent versus public code: where the sources differ

The headline claim in coverage is proactivity. DeepSeek says that if someone reports a bug in a group chat, the agent can identify the issue, try to reproduce it and submit a pull request if it finds a fix, without anyone tagging it. DeepSeek says an internal Super-agent system is already running, reproducing and fixing software problems, responding to code reviews and improving DeepSeek Harness itself. It also says the internal version monitors public developer discussions for new information.

The public prototype is described more cautiously. Pillitteri’s review says the published agent acts only when an authorized person asks it to or when a rule written by the team allows it. The README states that the published code does not match the internal version or the one in the paper, and that some ideas in the paper remain future directions that are not implemented. Global Times likewise reports that the public version differs from the internal system and that some features are still under development.

Builders should read “starts work without being asked” as a description of DeepSeek’s internal system and design goals. In the released code, unprompted action is gated by team-written rules, according to the repository review.

An agent designed to stay quiet

The project’s philosophy document sets the tone. Its principle is to “take as little human attention as possible,” balanced against delivering as much value as possible; DeepSeek’s user manual on GitHub phrases the goal as “take as little human attention as possible, and give as much value as possible”. Every message, notification or panel line is treated as a cost to someone’s attention that has to be worth more than it costs.

Access: what you need to run dsh-superagent

The code is public and free under the MIT licence. It installs into a dedicated DeepSeek Harness profile and needs version 0.2.1-alpha.2 or later of the program. The listed requirements are:

• A dedicated DeepSeek Harness profile
• A GitHub account reserved for the agent
• A configured model provider
• Node.js 22.19 or later, pnpm, git, tmux, Python 3.8 and GitHub CLI

Each installation works on a single repository. The repository was created on October 31 with a single commit titled “add the dsh-superagent research prototype,” and at 7:27 UTC on November 1 it had 4,137 stars, 381 forks, 24 watchers and no open issues. The README cites the paper but does not link to it.

What it means for builders, and the open risks

For teams building on coding agents, the release is mainly a design reference. It shows one concrete way to separate long-lived context (the Router) from disposable execution (Workers), to scope each task with an explicit Mandate that cannot grow, and to keep long test runs from blocking a conversation. These patterns can be studied regardless of which model sits underneath, since the prototype needs a configured model provider.

Analyst Ma Jihua told Global Times that the work reflects a shift in agents toward parallel processing, event-driven responses and greater autonomy. He said:

“Users no longer judge AI agents solely by their models. Task success rates, the ability to recover from failures and the cost of managing context are becoming increasingly important.”

The sources raise clear limits:

• No independent evidence. No source reports third-party benchmarks or tests. Performance claims come from DeepSeek, and the reviewer of the code did not run it.
• Reliability of unprompted work. Trending on Weibo notes that an agent which acts without being asked “can also be wrong without being asked,” and that putting humans in charge of acceptance assumes the submitted work is worth accepting.
• Access and permissions. The agent needs its own GitHub account and access to team chats, so teams should decide carefully which repositories and channels it can see.
• Not production software. DeepSeek says the prototype is not an official product or production-ready.
• Unsolved system problems. Global Times notes remaining challenges in scalability, long-term memory and cross-agent communication.

The competitive context around the launch

The release lands as Chinese firms race US developers to build more advanced and affordable AI tools, with much of the attention shifting to agents that automate work for developers. DeepSeek has raised about 100 billion yuan (S$19.2 billion) in its latest fundraising round ahead of a planned initial public offering, according to The Straits Times. Stocktwits ties the timing to Anthropic, which it reports filed its S-1 registration statement with the US Securities and Exchange Commission in late October and is expected to begin its IPO roadshow this month.

Other Chinese companies are pushing agent platforms too. Tencent said its Agent Development Platform lets enterprises customize, manage and coordinate agents, and Alibaba Cloud introduced AgentLoop and AgentTeams in July to monitor and coordinate multiple agents. Trending on Weibo frames DeepSeek’s move as competing on scaffolding rather than model benchmarks.

What to watch next

• Whether DeepSeek publishes the full paper in a form that can be checked against the code, since the README cites it without a link.
• How much of the internal system’s unprompted behaviour reaches the public prototype, given the stated gaps between versions.
• First hands-on reports from developers running the prototype, and whether issues and pull requests start appearing in a repository that so far has one commit.
• Whether the pattern holds up in practice; Trending on Weibo says that will become visible in the next model cycle.