mishal brands

Trend · AI agents

DeepSeek published a research coding agent. Here is who should wait

Updated Nov 2, 2026 · By Mishal Brands Editorial

Checked Nov 2, 2026. The description of the prototype comes from its GitHub repository: the README, the license, the setup guide, and the numbered docs. News context comes from The Next Web, the South China Morning Post, and MarkTechPost, all dated Nov 1, 2026. We have not installed or run this prototype.

The short answer

dsh-superagent is a research prototype from the DeepSeek Harness team, published in a public GitHub repository on Oct 31, 2026, under the MIT license. It watches one GitHub repository, and when a person the team has allowed asks for work, it opens a work item, does that work in its own session and its own folder, and replies in the same conversation, for example with a pull request.

What it does, step by step

  1. It watches one repository on github.com. The README says it is a coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. You can also turn on Feishu, a workplace chat app. The deployment guide says GitHub Enterprise Server is not supported.
  2. It does not start work for just anyone. An allowed list says who may ask. The philosophy note says it acts only when asked, or when a written rule the team wrote allows work that nobody asked for.
  3. An allowed person asks in the place the team already uses. On GitHub, that is a mention of the agent's account on an issue or in a comment. The setup guide's try-it step is a comment of that form.
  4. A router session takes the request, and a worker session does the job. The execution doc says the main session uses a preset named dsh-superagent-router, and each work item gets its own session, named for that item, plus its own folder. The deployment guide says those two presets are only for this bundle.
  5. It answers where the request was made. The README's example is a pull request. The philosophy note says it confirms once, stays quiet while it works, and writes again only when there is a result, a question, or a failure.
  6. A panel on that same computer lists what still needs a person. The panel doc says to open Superagent in the DeepSeek Harness sidebar. The deployment guide serves that panel on the machine itself, at 127.0.0.1. It is not a public website. The philosophy note says a healthy state should take no space on the panel.

Who should try it now

Developers who are comfortable running research code on a machine kept separate from everyday work. The setup guide is written so a coding agent walks through it with you, one question at a time, and installs only after you confirm the plan.

The setup guide and the deployment guide list what that machine needs: Node.js 22.19 or newer, or Node.js 24 or newer; pnpm; git 2.32 or newer; Python 3.8 or newer; tmux; GitHub CLI 2.48 or newer; and DeepSeek Harness with a model provider configured. The package file's engines field matches the Node.js range.

Who should wait

Everyone else. The README says this is a research prototype for reference only, not an official DeepSeek product, and not production-ready software. It says the code may still contain bugs, and that its style may be inconsistent. It also says the published code is not identical to the team's internal version, or to the design article Exploring Super-agent Design: Towards a More Autonomous Agent System. Some ideas in that article are not built in this prototype.

The catches

  • It is a prototype, not a product you subscribe to. There is no hosted app to open in a browser. You run the code yourself, on one repository.
  • It runs with the full access of the user account that starts it. The deployment guide says both presets use danger-full-access as that operating-system user. A worker can run programs and read that user's home folder. The guide says the account wrappers choose which GitHub login is used, and that they do not wall the process off. Its isolation advice is a dedicated operating-system account that holds only the agent's credentials.
  • Work that nobody asked for stays off until you turn it on. The authorization doc lists three accept modes: off, observe, and auto. Observe is the default. In observe, the agent records unasked work it would have accepted and does not do it. A direct request from an allowed person still runs, in every mode. The doc says to switch to auto only after those recorded decisions look right.
  • The code is free. A model still has to be configured, and this repository does not price it. The license file is the MIT License, copyright 2026 DeepSeek Harness Team. That is an ongoing free license to the code, not a free trial and not a one-time pile of credits. The deployment guide says dsh-superagent uses the DeepSeek Harness profile's model unless a setting names another. No price for that model appears in the repository, so this page does not invent one.
  • A news write-up describes a simulated speed test. Those figures are not in the repository, so they are not repeated here. MarkTechPost's Nov 1, 2026 article describes a timing comparison and says the test used synthetic messages and simulated tools, not real coding work. It also says the figures come from secondary reports of a design article. The README says the published code is not the same as that article.

How it compares with asking a coding assistant directly

When you ask a coding assistant in a chat, you stay in that conversation: you explain the job, you wait, and you check back. This prototype is built for a different habit. It listens on GitHub, and optionally on Feishu. An allowed person asks in the place the team already talks. The agent works in its own session and answers there, for example with a pull request, and a panel shows only what still needs a person. The README's line for that habit is to take as little human attention as possible. News coverage on Nov 1, 2026 placed the idea next to coding assistants from other labs. The Next Web, citing Bloomberg, said it is aimed at Anthropic's Claude Code. The South China Morning Post described a coding-assistant race with OpenAI and Anthropic, and said the aim is to cut the work of coordinating several agents and watching their progress. The repository itself does not name those products.

Related on Mishal Brands

Sources (checked Nov 2, 2026)

  • GitHub, deepseek-ai/dsh-superagent-prototype, README: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/README.md
  • The repository's commit on Oct 31, 2026, "feat: add the dsh-superagent research prototype": https://github.com/deepseek-ai/dsh-superagent-prototype/commit/6337fc823ce20251c8de8085948165c81aaa701e
  • MIT license, copyright 2026 DeepSeek Harness Team: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/LICENSE
  • docs/000-philosophy.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/000-philosophy.md
  • docs/001-domain.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/001-domain.md
  • docs/002-authorization.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/002-authorization.md
  • docs/003-execution.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/003-execution.md
  • docs/005-panel.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/005-panel.md
  • docs/006-deployment.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/docs/006-deployment.md
  • SETUP.md: https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/SETUP.md
  • package.json (Node.js engines): https://github.com/deepseek-ai/dsh-superagent-prototype/blob/main/package.json
  • MarkTechPost, "DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent" (Nov 1, 2026): https://www.marktechpost.com/2026/11/01/deepseek-dsh-superagent-router-worker-coding-agent/
  • The Next Web, "DeepSeek open-sources Super-Agent, a coding agent that reads team chats" (Nov 1, 2026): https://thenextweb.com/news/deepseek-super-agent-open-source-coding-agent
  • South China Morning Post, "DeepSeek unveils AI 'super agent' in coding race with OpenAI, Anthropic" (Nov 1, 2026): https://www.scmp.com/tech/article/3372884/deepseek-unveils-ai-super-agent-coding-race-openai-anthropic

Disclosure: this page has no affiliate links. No company paid for or reviewed this page. See our affiliate disclosure and how we review.

Read next: All trends · dsh-superagent in the library

Browse by category: All 31 Development tools. Every tool page lists the vendor’s own price and the date we checked it.