DeepSeek Harness, the open-source agent that Super-agent is built on. Image: DeepSeek
DeepSeek wants its AI to stop waiting for instructions. The Chinese lab’s Harness team has open-sourced dsh-superagent, a prototype coding agent that sits in a team’s GitHub repository and work chat and can pick up a problem, and propose a fix, before anyone asks it to.
The code, released under the MIT licence, comes with a long essay the team posted on its official WeChat account on the evening of October 31, Beijing time, titled “Super-agent Design Exploration: Towards More Autonomous Agent Systems”. The essay says DeepSeek already runs its own Super-agent internally, where it reproduces and fixes bugs, responds to code reviews and helps build DeepSeek Harness itself.
An agent that steps in on its own #
Most coding agents, including Anthropic’s Claude Code and OpenAI’s Codex, work on request: you open a session, give it a task and watch it run. Claude Code’s GitHub integration, for example, acts when someone mentions @claude in an issue or pull request.
The essay describes something different. DeepSeek’s internal instance, it says, keeps reading feedback in public group chats, works out the root cause of a reported problem and submits a pull request with a fix, without anyone tagging it.
The team argues that agents are now held back less by models than by people. In its experience a person can usually keep track of only three to five active agent sessions at once, so how much work runs in parallel is capped by human attention rather than by what the model can do.
How Super-agent works #
Its answer is a two-layer design. A permanent “Router” handles every conversation with people and hands out jobs but never does the work itself; the essay says a lighter, faster model is often enough for it. Separate “Worker” sessions, each with its own context, carry out tasks in parallel. The design also keeps the agent’s computing, memory and working environment apart, and adds an outside “Watchdog” that recovers from failures, so the agent keeps running when a developer’s own computer goes offline.
The released prototype is narrower: a “proactive coding agent for one GitHub repository”, packaged as a bundle for DeepSeek Harness, the lab’s free, open-source rival to Claude Code and Codex. It listens to GitHub and, optionally, Feishu, ByteDance’s workplace chat app (sold as Lark outside China). When someone asks it for work, it opens a ticket, works in its own session and folder, and replies where the request was made, for example with a pull request. A web panel lists only what needs a person.
Proactive, but on a leash #
Unprompted work is off by default. The project’s documentation says the agent “acts only when asked, or when a written rule allows it”, and that work nobody asked for “starts in a mode that only records what the agent would have done”. Teams write rules describing the jobs it may take on. In the default “observe” mode it only logs what it would have accepted; it acts on its own only once a team switches to “auto”. It can read messages that don’t mention it only where a team lets it receive everything in a chat or repository.
The project’s guiding principle is to “take as little human attention as possible, and give as much value as possible”. The docs remind operators that the agent posts “under accounts your team sees every day”, tell them to start with a short list of people allowed to give it work, and add: “Do not use it to produce volume.”
Not a product, DeepSeek says #
DeepSeek is careful about what this is. The README calls it a “research prototype for reference only, not an official DeepSeek product”, says much of the engineering was done by coding agents, and warns that it is “research code, not production-ready software” that may still contain bugs. It also differs from DeepSeek’s internal system: “Some designs discussed in the article are directions for future development and are not implemented in this prototype.”
It caps a busy few weeks for the Hangzhou lab, which has open-sourced the inference engine behind its own service, raised about $15 billion and upgraded its cheap V4.1 Flash model, which it says beats Claude Opus 5 at some coding tasks. OpenAI is chasing a similar idea for consumers: at DevDay in September it unveiled always-on “dots” agents that keep working between conversations.
Why it matters #
Coding agents have mostly waited for a prompt; DeepSeek is betting the next step is one that listens in and acts on its own, within rules a team sets. By publishing both the blueprint and working code under the MIT licence, it hands every developer a free starting point to compare with Anthropic’s and OpenAI’s paid tools.
Sources: deepseek-ai/dsh-superagent-prototype on GitHub (README, documentation and MIT licence); DeepSeek Harness team, “Super-agent Design Exploration: Towards More Autonomous Agent Systems” (official WeChat account, Oct 31, 2026); Claude Code GitHub Actions documentation.