DeepSeek's Harness team published a design for a "super agent" on Oct. 31 and released a working prototype, dsh-superagent-prototype, under an MIT license. The coding agent can follow signals in group chats and open pull requests without anyone pinging it, according to IT之家.
The README is careful about what this is not. It calls the code a "research prototype for reference only, not an official DeepSeek product," and the company has offered no general availability date. Still, the release hints at where DeepSeek wants to take agents after its reported $15 billion fundraising round.
What the agent does
The public code is a repository-scoped, proactive coding agent bundle for DeepSeek Harness, the South China Morning Post reported Nov. 1. It listens to GitHub and, optionally, Feishu, the workplace chat platform. From there it opens tickets, runs each in its own working directory and session, and replies in the original thread, often in the form of a pull request. A panel flags "what needs a person."
IT之家 described the central use case as coding workflows. The agent watches public group-chat signals, tries to reproduce a reported bug, analyzes the cause and opens a pull request, all without a direct request.
The Harness team wrote on X on Nov. 1 that its internal super-agent instance "is already running continuously." The public repository differs from that internal version, the team said, and some designs in the write-up are future directions that have not been implemented.
How it is built
According to IT之家, the design splits work into two layers. A Router handles human interaction and global context. Workers carry out tasks in isolated sessions and report back to the Router, which remains the single outward-facing voice.
External events, such as instant messages, calendar items, code review comments and test results from continuous integration, become "Triggers." Programmatic filtering strips out noise before the model decides what to do. Workers can subscribe to a Thread so that related events reach them directly, which eases the load on the Router.
Other components include a watchdog session that monitors for infrastructure errors and for stretches with no real progress. An asynchronous loop lets the chat continue while slow, expensive tool calls run in the background. The agent loop and its state live in the cloud, though execution can stay on a user's own machines through a lightweight daemon.
