DeepSeek Publishes Super-Agent, a Coding AI That Can Work Without Direct Requests

DeepSeek published the source code for Super-Agent, an experimental coding agent built to handle tasks under team-defined rules and, if permitted, take on unrequested work. The prototype is designed to work with one GitHub repository, listen to GitHub and optionally connect to Feishu, where it can receive requests and send results back to the original conversation.

That autonomy comes with limits: unrequested work requires a rule written by the team and is disabled by default. In that mode, Super-Agent only records what it would have done rather than carrying out the task. DeepSeek's authorization documentation details how those rules work.

When an authorized person assigns a task, the agent opens a work item and handles it in its own session and directory. It confirms the request once, then stays quiet until it has a result, needs a decision or cannot finish. A web panel collects situations that need human attention.

The prototype is packaged for DeepSeek Harness, the company's environment for running agents. DeepSeek says Super-Agent is research code, not an official product, and warns that it may contain bugs and inconsistent engineering practices. The public version differs from the company's internal one, and some designs described in related material are not implemented yet.

DeepSeek also says coding agents did much of the engineering work on the project. The Next Web describes Super-Agent as a direct competitor to Anthropic's Claude Code.

How much freedom should coding agents get to act on work nobody directly assigned?