DeepSeek Super-agent: China's open-source coding agent that fixes bugs unasked

DeepSeek, the Chinese AI lab, has released a prototype called Super-agent: a coding agent that sits in a team’s GitHub and chat channels, picks up bugs and requests, and works on them without waiting for someone to type a command. The company’s Harness team announced it on WeChat on Saturday, 31 October, and put the code on GitHub under the MIT licence, so anyone can download, study and change it for free.

It is being read as DeepSeek’s answer to Anthropic’s Claude Code and the other paid coding agents that have taken over a lot of software work this year.

What Super-agent does

Most coding assistants wait for a prompt. Super-agent is built to watch instead. According to the project’s own documentation, it monitors one GitHub repository and, optionally, the team’s Feishu chat (a Chinese workplace app). When an authorised person asks for something, or a situation matches rules the team has set, it opens a work item and handles it in its own separate session and folder.

The design has two parts. A Router decides which requests to take on. Isolated Workers then do the job: reproducing a bug, writing the fix, answering code-review comments. The result comes back to the original conversation, often as a pull request a human can approve. A web panel lists anything that needs a person’s decision.

DeepSeek describes the goal plainly: use as little of the developer’s attention as possible while doing as much useful work as possible.

How good is it?

DeepSeek says an internal version is already part of its own daily research and development work, including reproducing issues and responding to code reviews. In its write-up, the team reported early test results showing tasks finishing faster than with a conventional setup. Those are the company’s own numbers and have not been checked independently.

The company is also unusually blunt about the limits. The GitHub page calls it a research prototype and “not an official product”, says much of the code was itself written by coding agents and may contain bugs, warns it is not ready for production, and says the public version differs from what DeepSeek runs internally. Some ideas in the accompanying article are not built yet.

It runs on DeepSeek Harness, the open agent toolkit the company released in August, which DeepSeek itself labels experimental software without a security audit.

Why it matters

The race in AI has moved from chatbots to agents that do the work, and coding is where the money is. US companies sell these tools on monthly subscriptions. A Chinese lab giving away the blueprint for a proactive agent puts price pressure on them, the same way DeepSeek’s cheap models did last year. It also lands as DeepSeek reportedly raises around $15 billion ahead of a planned share listing, according to media reports.

What it means for Pakistan

Pakistan’s freelancers and software houses live on exactly the kind of work these agents now automate: bug fixes, small features, code reviews. A free, open tool that any team can run on its own server lowers the cost of trying agents without paying in dollars every month. It also raises the bar: clients will increasingly expect developers who can manage an agent’s output, not just write code by hand.

One caution for anyone tempted to try it today: an agent that can open pull requests on its own needs careful permissions. DeepSeek itself says its sandbox does not guarantee isolation, so test it on a throwaway repository first, not on a client’s live code.

For now, Super-agent works only with GitHub and Feishu. Whether DeepSeek adds Slack or other tools, or turns the prototype into a supported product, is the next thing to watch.