DeepSeek open-sources a prototype agent that ships PRs

Most coding agents wait in your terminal. DeepSeek just published one that waits in your repo and answers with a pull request.

DeepSeek put dsh-superagent on GitHub under the MIT license. The README calls it a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness, the company's agent harness.

It takes requests in the repo and answers with a pull request

The agent listens to a GitHub repository and, optionally, to Feishu, the workplace chat app.

When an allowed person asks for work, it opens a work item, does the job in its own agent session and directory, and replies in the conversation where it was asked. The README's example reply is a pull request. A web panel lists the items that need a person.

Per the README, the agent acts only when someone asks or when a rule the team wrote allows it. It speaks only when someone must act or can use a result. The stated goal is to take as little human attention as possible.

Setup is meant to be done by another coding agent. The README tells you to ask yours to read SETUP.md and set the system up with you. That agent asks one question at a time, shows the full plan, then installs and configures everything.

The repo ships nine numbered design documents, from philosophy and authorization to deployment, architecture and a cookbook.

Why it matters: the agent moves to where the team already talks

Yes, it's a prototype, and DeepSeek says so plainly. The README calls it "Research prototype for reference only, not an official DeepSeek product."

The design is still worth your time. Most agentic coding tools sit in one developer's terminal and wait for that developer. This one sits in the repository and the team chat, and anyone on an allow list can hand it work.

That moves the hard question. It stops being "can the model write the fix" and becomes "who is allowed to ask, and what may it do without asking." DeepSeek gave authorization its own design document. That tells you where it thinks the risk is.

What to know before you try it

Read the limits first. The README warns the code "may still contain bugs," that its style is inconsistent, and that you should review and adapt it before using it in your own environment.

It also says the public version differs from DeepSeek's internal one, and from the article it accompanies, "Exploring Super-agent Design: Towards a More Autonomous Agent System." Some designs in that article are future directions the prototype doesn't implement.

There are no benchmark results in the README, and it doesn't name a model.

The MIT license means you can fork it, change it and ship it commercially. If your team builds internal agents, the authorization and execution documents are the parts to borrow even if you never run the code.

One caution for some readers. If you work in or for a US government agency, TopFive reported that DHS banned DeepSeek, and tools built on it, from agencies. Check that rule before anyone installs this.

The repo had more than 4,000 stars when TopFive checked, with a single commit in its history. People are reading it.

What to watch

Watch whether DeepSeek turns this into a supported product inside DeepSeek Harness, and whether the article's unbuilt designs show up in the code.

Then read document 002, the authorization one, before you let any agent take requests from your team.