DeepSeek releases public research prototype of its Super-agent

DeepSeek has released a public research prototype of its Super-agent, the agent system it built for its own internal use. The release targets a specific workflow. A team points the agent at one GitHub repository, and the agent watches for events that match rules the team has written. When it spots a matching event and someone allowed asks for work, it opens a work item, runs it in its own agent session and directory, and replies in the original conversation, for example with a pull request. The Turing Post newsletter, which covered the release in a short item, described the prototype as open-sourced and gave no license or performance details. The project's GitHub page, however, lists an MIT license and 18.4k stars.

The mechanics matter most. According to the repository's README, the prototype listens to GitHub and, optionally, to Feishu messages. The operator can allow certain people to ask the agent for work. The agent acts only when someone allowed asks or when a rule the team wrote permits it. The README frames this as a design principle: the agent should take as little human attention as possible while giving as much value as possible, and it speaks only when someone must act or can use a result. A web panel shows which items need a person. The README warns that an agent able to act for a whole team can also waste the team's attention.

The README also qualifies the release in ways the announcement does not. It calls the code a research prototype, "not an official DeepSeek product," and says it was adapted for publication. As a result, it is not identical to the internal version or to the implementation described in an accompanying article titled Exploring Super-agent Design. Some designs in that article are future directions that the prototype does not implement. The README also says much of the engineering was done by coding agents, and that the code may contain bugs and inconsistent practices. It tells users to review and adapt the code before running it in their own environment.

Setup is itself delegated to an agent. The README tells users to ask a coding agent to read SETUP.md and configure the system with them. According to the README, the agent asks one short question at a time before installing everything. The repository also lists documentation in numbered order, covering authorization, execution, prompts, a panel, deployment and architecture. It names a single check command, pnpm check, for typecheck, tests and build.