DeepSeek open-sources a coding agent built to bother you as little as possible

DeepSeek's Harness team on Saturday, Oct. 31, 2026, published dsh-superagent-prototype, an MIT-licensed "proactive coding agent for one GitHub repository," on GitHub alongside a design article, "Exploring Super-agent Design: Towards a More Autonomous Agent System." The repo passed 6,000 stars within a day.

Every coding agent pitch is about doing more. DeepSeek's is about interrupting you less, which is a sneakier and possibly smarter bet: once agents can all write a decent pull request, the scarce resource is the human who has to read the notifications. Caveats are loud, and DeepSeek wrote them itself: it's a prototype, the speedup comes from a simulation, and nobody measured whether the code is any good. But open-sourcing the plumbing for a team-wide agent, MIT-licensed, the same week rivals are charging for theirs, is a very DeepSeek move.

The README opens: "Research prototype for reference only, not an official DeepSeek product." It says much of the engineering was done by coding agents, that it may contain bugs, and that the published version is not identical to DeepSeek's internal version or to everything described in the design article. Those lines are the repo's. This release is not a new DeepSeek model.

What it is: a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. A harness, here, is the software wrapped around a model so the model can take steps, not the model itself. It listens to GitHub and, optionally, Feishu, a workplace messaging app. When an allowed person asks for work, it opens a work item, which it calls a Ticket, works on it in its own agent session and its own directory, and answers in the original conversation, for example with a pull request. A web panel shows only what needs a person. Those details are the README's.

Its stated rule: "take as little human attention as possible, and give as much value as possible." It acts only when asked, or when a rule the team wrote allows it, and it speaks only when someone must act or can use a result. That rule is DeepSeek's.

The design article, as MarkTechPost reported it, splits the system in two. A long-running Router handles interaction and dispatch. Isolated Worker sessions do the actual work. The article argues a person can actively follow only about 3 to 5 agent sessions at once, so human attention, not model capability, becomes the bottleneck. Those points are the design article's, via MarkTechPost. The README says the published code is not the same as everything the article describes.

Release details from the repo: TypeScript, v0.1.0, one commit, MIT license. It uses whatever model the Harness profile is set to. It runs on your own machine and needs Node.js, git, Python, tmux, and the GitHub CLI. Tmux is a tool that keeps terminal sessions running in the background. MarkTechPost notes it runs with the full permissions of the user account on that machine, so it should be isolated before use. Those lines are the repo's and MarkTechPost's. None of the sources compare its coding quality with Claude Code or with OpenAI's Codex.

DeepSeek's own simulation used synthetic message streams of about 300 messages, with about 20 real requests, and simulated tools. It reported a p95 task-completion time of 15.7 minutes, versus 29.5 minutes for a single-session baseline, about 1.9 times faster. p95 means 95 out of 100 of those simulated tasks finished within the stated time. Real coding quality was not measured. All of those figures are DeepSeek's, from the simulation.

The repo was created on Oct. 31. By Sunday morning, Nov. 1, it had about 6,200 stars and about 600 forks. Bloomberg, in "DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic," and the South China Morning Post, in "DeepSeek unveils AI 'super agent' in coding race with OpenAI, Anthropic," both on Nov. 1, framed the release as DeepSeek entering the autonomous coding-agent race against Anthropic's Claude Code and OpenAI. The star and fork counts are the repo's that morning. The race framing is those two reports'.

The picture is a screenshot of the GitHub page for deepseek-ai/dsh-superagent-prototype. It is a public repo. The about line reads "Towards a More Autonomous Agent System." The page shows 6.2k stars, 596 forks, an MIT license, and one commit by the DeepSeek Harness Team, "feat: add the dsh-superagent research prototype," with folders including docs, examples, prompts, skills, src, and tests. The credit is DeepSeek Harness team / GitHub (screenshot).

In plain terms, DeepSeek published research code for a coding agent that watches one GitHub repository, splits the work across separate sessions, and tries to interrupt a person only when a person has to act. The company says it is not a product, the published code is not the internal version, and the speedup comes from a simulation that did not grade whether the code was any good. News coverage put the release in the same race as Claude Code. Nothing in the sources says it writes better code than those tools.