DeepSeek Open-Sources Team Coding Agent

DeepSeek on Oct. 31 open-sourced dsh-superagent, a prototype coding agent that takes requests from a team's GitHub repository and, optionally, its Feishu chat, then answers with work such as pull requests. The Chinese AI lab released the code under the MIT license and calls it "a research prototype for reference only, not an official DeepSeek product."

Key takeaways

• DeepSeek's Harness team published dsh-superagent on GitHub on Oct. 31 under the MIT license, as research code rather than a supported product.
• The agent works on one GitHub repository, acts only on requests from allowed people or team-written rules, and replies in the original thread, for example with a pull request.
• Its design rule is to "take as little human attention as possible, and give as much value as possible," and unrequested work starts in a record-only mode.

What DeepSeek's dsh-superagent does

According to the project README, dsh-superagent is a proactive coding agent for a single GitHub repository, packaged for DeepSeek Harness, the company's open-source agent app. It listens to GitHub events and, if configured, to Feishu, the workplace chat app from ByteDance.

When an allowed person asks for work, the agent opens a work item and handles it in its own agent session and directory. It then answers in the conversation where the request came from. A web panel lists only the items that need a person to act.

When AI Newsway checked the repository on Nov. 2, it showed about 3,300 stars and 293 forks.

An AI coding agent built around human attention

The project's philosophy document judges every message, notification and setting by the attention it costs. A request counts only when it addresses the agent directly and comes from an allowed person or a conversation the team has opened to requests.

The agent does not start unrequested work on its own. That requires a rule the team writes, and such work begins in a record-only mode where the agent logs what it would have done.

The document also puts responsibility on operators. "Every message it sends costs your colleagues' attention, and you are responsible for it," it says, and it advises against using the agent for volume such as mass pull requests.

Limits of the open-source release

DeepSeek tells users to "treat it as research code, not production-ready software." The README says the published code is not identical to DeepSeek's internal version or to the design article that describes it, and that some of those designs are not built yet.

Context: DeepSeek takes aim at Claude Code

Bloomberg reported on Nov. 1 that the release is aimed at rivaling Anthropic's Claude Code. It plugs into DeepSeek Harness, an open-source agent app the company says is in public preview worldwide.

Team-level agents face a known hurdle. CooperBench, a benchmark of more than 600 collaborative coding tasks posted to arXiv in January, found agents averaged about 30% lower success rates when working together than when doing the same tasks individually. That study did not test DeepSeek's agent.

Earlier research on LLM bug-fixers that broke working code points the same way: agent output still needs human review.

What's next

DeepSeek has not said whether dsh-superagent will become a product. Its guidance to operators is to start with a short allowed list and record-only proactive work, then widen both once results look right. For U.S. public-sector users, DeepSeek products remain under a Department of Homeland Security removal order for federal civilian agencies published Oct. 20.

FAQ

Is DeepSeek's super-agent free to use?

Yes. The code is released under the MIT license, which allows commercial use and modification. DeepSeek says it is a research prototype, not an official product.

Does dsh-superagent work without Feishu?

Yes. GitHub is the required event source, and Feishu is optional.

Will the agent make changes nobody asked for?

Not by default. Unrequested work needs a team-written rule and starts in record-only mode.

References

• deepseek-ai/dsh-superagent-prototype — GitHub, DeepSeek (Oct. 31, 2026)
• 000-philosophy.md — GitHub, DeepSeek
• DeepSeek Harness — DeepSeek
• CooperBench: Why Coding Agents Cannot be Your Teammates Yet — arXiv (Jan. 19, 2026)
• DeepSeek open-sources Super-Agent, a coding agent that reads team chats — The Next Web (Nov. 1, 2026)