DeepSeek releases free agent that monitors GitHub and acts only when an approved person asks

What's new

DeepSeek has released Super-agent, which is available on GitHub under the name dsh-superagent. It is a coding agent, meaning a program that writes and fixes code for you. According to Complete AI Training, it was announced on a Saturday by DeepSeek's Harness team in a post on WeChat. The exact date has not been confirmed.

What sets it apart from many other tools is that it works proactively rather than just waiting for you to ask it to do something. According to the README, it is built for a single GitHub repository (the place where a project's code lives), and it monitors that repository and, optionally, Feishu group chats. According to DeepSeek, it only opens a task when an approved person asks for work or when a rule written by the team allows it. It then works in its own session and replies in the original conversation, for example with a pull request (a proposal for code changes).

The code is open source under the MIT license, so anyone can download it, use it and build on it. It is packaged as a solution for DeepSeek Harness, the company's platform for agents.

What's smart

Under the hood is a two-layer design. A permanent "Router" handles communication, while separate "Worker" sessions carry out tasks side by side. A "Watchdog" process keeps an eye out for errors and for agents stuck in a loop.

According to DeepSeek, the human's role is deliberately kept small. A local web panel shows only what requires a human decision. The agent also starts in a mode where it only logs what it would do before it actually acts. That makes it possible to see how it behaves first. DeepSeek's own design principle is to demand as little human attention as possible and deliver as much value as possible.

Digital Trends explains that the tool is built on "agent harness" technology, meaning software wrapped around a language model that enables it to handle complex tasks.

Cheaper or better?

The tool itself is free to download, but no price is given, and the sources include no benchmarks or speed figures. You have to set it up yourself and arrange access to a language model yourself, so costs may arise there. The sources do not say whether it can be used in Denmark and the EU, beyond the fact that it runs through GitHub.com.

Digital Trends and Complete AI Training see it as a challenger to Anthropic's Claude Code, and Digital Trends also mentions OpenAI's Codex. There is no direct comparison on price or quality. Digital Trends assesses that the prototype is not yet proven for complex large-scale deployments.

Important caveat: the README file calls the project a research prototype for reference only and not an official DeepSeek product. Large parts were written by coding agents and may contain errors. The public version differs from the internal one, and some designs are only future directions.

Who it's good for

The agent is aimed at developers and technical users who want to experiment with a coding agent that looks after a project on its own. The requirements are a dedicated GitHub user account, Node.js 22.19 or newer and GitHub CLI 2.48 or newer. It only works with GitHub.com, not Enterprise Server, and a separate operating system account is recommended for security reasons.

For ordinary users and most companies, it is hardly ready for daily use. Interest is high, though: at the time of checking, the repository had more than 11,600 stars and more than 1,200 forks. Both outlets also note that, according to Bloomberg, DeepSeek has raised around 100 billion yuan (about 15 billion dollars) ahead of a planned stock market listing.