DeepSeek Super-Agent: Open Coding Agent, What to Check

On October 31, 2026, DeepSeek published the code for Super-Agent on GitHub: a coding agent that works on one repository and takes requests from GitHub and from the Feishu team chat. It is MIT-licensed but labelled a research prototype. It matters to anyone who lets an AI agent change their website's code.

What we know

What it is: according to the README of the «dsh-superagent-prototype» repository (October 31), a "proactive" coding agent for a single GitHub repository, packaged for DeepSeek Harness, DeepSeek's open-source agent app, currently in public preview.
How it works: it listens to GitHub and, optionally, Feishu, ByteDance's workplace chat. When an allowed person asks for work, it opens a work item, handles it in its own agent session and directory, and replies in the original conversation, for example with a pull request. A web panel shows what needs a human.
Its core rule: the repository's philosophy document says the agent acts only when asked or when a rule the team wrote allows it, and that unrequested work starts in a mode that only records what it would have done.
What it is not: the README calls it a research prototype, "not an official DeepSeek product". It warns of possible bugs, says it is not production-ready, and notes the public version differs from DeepSeek's internal one. Some ideas from its design article are not implemented.
How the press frames it: Bloomberg, as cited by The Next Web, says DeepSeek is aiming it at Anthropic's Claude Code and that it can watch group chats and suggest fixes without being summoned. That goes further than the philosophy document, which requires a written rule and starts in record-only mode.

What changes and what doesn't

What changes: a major AI lab has open-sourced a design for an agent that works on behalf of a whole team, not just the person who launched it.
What doesn't: nothing for non-developers. It is not a consumer app, has no published price and takes several technical steps to set up.
Still true: whatever an agent does in the team's name, the team owns.

How to tell if it affects you

If a team or vendor maintains your website with coding agents (this one or any other), ask and check:

Which repository and permissions the agent has. Minimum: one repository, permission to propose changes, not to deploy them.
Who can give it work. The prototype itself relies on a short list of allowed people; ask for that list to exist in writing.
What it does unprompted. The sensible default is the one in the philosophy document: record only first, widen once the recorded actions look right.
That every change is reviewed. An agent's pull request should be reviewed and tested before it reaches your live site, like a person's.
What sits in the chats it reads. If the agent listens to team chat, no passwords, API keys or customer data belong there.
Prototypes stay off production. The README itself asks users to review and adapt the code before using it.

Sources

Updates: if DeepSeek turns it into a product or publishes pricing, it will be added here with a link.