DeepSeek open-sources dsh-superagent, a proactive coding agent that works GitHub requests

What DeepSeek published

DeepSeek published dsh-superagent in its deepseek-ai/dsh-superagent-prototype repository on 31 October 2026, with the commit signed by the DeepSeek Harness Team. The code is under the MIT license. The repository accompanies a design article titled Exploring Super-agent Design: Towards a More Autonomous Agent System, and the README warns that some ideas in that article are future directions that the prototype does not implement.

The README describes dsh-superagent as a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness, the company's agent runtime. It listens to GitHub and, if configured, to Feishu, the workplace chat app. When a permitted person asks the agent for work, it opens a work item, handles it in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows only the items that need a person.

DeepSeek is blunt about the code's status. 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 may contain bugs and uneven style. The published version was adapted for release and differs from the one the team uses internally. Bloomberg reported on 1 November that the release is aimed at challenging Anthropic's Claude Code, and that the Harness team was set up earlier this year to build agents that compete with those from US labs. The repository had more than 4,000 GitHub stars by the morning of 1 November.

How the agent decides when to act

The design rests on one principle, stated in the project's philosophy document: take as little human attention as possible, and give as much value as possible. In practice the agent acts only when someone asks it directly or when a rule the team wrote allows it. A request counts only if it addresses the agent and comes from a person on an allowed list, or from anyone in a conversation the team opened to requests. Messages, comments and quoted documents cannot change those permissions.

Work that nobody asked for needs an explicit rule, and it starts in an observe mode that records what the agent would have done without doing it. The agent confirms a request once, stays quiet while it works and writes again only when a result is ready, a decision is needed or it cannot finish. Each accepted request becomes a ticket with its own worker session. The documentation states that these worker sessions run commands without approval as the operator's operating system user, so the setup guide recommends a dedicated GitHub account for the agent, a short allowed list at first and no auto-merge of the agent's pull requests until people trust its output.

Setup is meant to be done by another coding agent: the README tells users to ask their agent to read SETUP.md and configure the tool with them, one question at a time. The prerequisites include Node.js 22.19 or later, pnpm, git, tmux, Python 3.8, the GitHub CLI and a compatible DeepSeek Harness build.

Each worker has limits that keep it responsive. Long shell commands move to background jobs after five seconds, so the worker can read new messages at its next step, and background work gets at most 30 minutes after a turn before the jobs are stopped and the ticket is reported as waiting. A person can stop any ticket, and a stopped ticket is not resumed by later messages unless the team accepts the work again.

What it means for developers

Most coding agents, including Claude Code, OpenAI's Codex and GitHub's agentic workflows, wait for a person to start a task in a terminal, editor or issue. DeepSeek's prototype treats the agent more like a team member that watches a repository's conversations, takes requests from several people at once and keeps parallel tickets open. The emphasis on cutting noise is a response to a real problem: an agent acting for a whole team can waste the whole team's attention. Because the code is MIT-licensed, other teams can study or adapt the routing and permission model, and early community ports to other agent runtimes appeared within hours.

The documentation is also written to be extended. Numbered documents cover the domain model, authorization, execution, the panel, deployment and architecture, and a cookbook explains how to add a new Source, the project's term for a stream of requests such as GitHub or Feishu. That makes the permission model, not the GitHub integration, the most reusable part of the release for teams that run their work in other chat or ticketing tools.

The open questions are practical. DeepSeek has not published benchmarks, the README does not name a default model, and the company has not said whether a supported product will follow. Teams trying it should treat it as research code that runs with broad local permissions. Earlier coverage of DeepSeek-V4.1-Flash-1017 noted that model was tuned for DeepSeek Harness. More agent news is in our AI agents section.

Questions people ask

What is DeepSeek's dsh-superagent?

A research prototype of a proactive coding agent for one GitHub repository, published by DeepSeek's Harness team on 31 October 2026 under the MIT license. It takes requests from GitHub or Feishu and answers with results such as pull requests.

Does dsh-superagent act without being asked?

Only under rules a team writes. By default it acts on direct requests from allowed people, and proactive work starts in an observe mode that only records what it would do.

Is it ready for production use?

DeepSeek says no. The README calls it research code that may contain bugs, and its worker sessions run commands without approval as the operator's user.

Sources

1. dsh-superagent-prototype: Towards a More Autonomous Agent System (DeepSeek, 31 Oct 2026)
2. 000 Philosophy (DeepSeek, 31 Oct 2026)
3. DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic (Bloomberg, 1 Nov 2026)
4. DeepSeek Expands AI Agent Ambitions With New Coding Prototype (Business Today, 1 Nov 2026)