By Hamza Faisal
DeepSeek wants its coding agents to stop waiting for instructions. The Chinese lab's Harness team has released dsh-superagent-prototype, an open-source coding agent that sits inside a team's GitHub repository and work chat, watches for problems, and proposes fixes before anyone asks it to. Published on October 31 under the MIT licence alongside a design essay on more autonomous agent systems, it argues that today's agents demand too much human attention.
The prototype is a bundle for DeepSeek Harness, the lab's free, open-source agent runtime and a rival to tools like Claude Code and Codex. Built for a single GitHub repository, it listens to repository events and, optionally, to Feishu, ByteDance's workplace chat app. When a permitted request arrives, it opens a work item, works on it in its own session and folder, and reports back where the request began — often as a pull request for the team to review and merge.
A router that delegates, workers that do the work
The design splits the job across two layers. A permanent Router handles every conversation with people and hands out jobs, but never does the work itself; DeepSeek says a lighter, faster model is often enough for that role. Isolated Worker sessions then carry out the actual coding, each in its own context, so several jobs can run in parallel without crowding a single conversation. The point: a person can actively follow only three to five sessions at once.
DeepSeek says its own internal Super-agent already goes further, reproducing and fixing bugs and responding to code reviews — reading feedback in group chats, working out the root cause of a reported problem, and submitting a pull request without anyone tagging it. For work the agent infers rather than receives explicitly, the public prototype defaults to an observe mode: it records what it would do without accepting the task.
Faster on paper, unproven in practice
In DeepSeek's own tests, the design finished tasks considerably faster than a conventional single-session agent under a noisy stream of messages. Each simulated run contained about 300 messages, of which roughly 20 were real work requests, and each setup was run five times. The super-agent approach reached a median completion time of 12.4 minutes and a 95th-percentile time of 15.7 minutes, against 29.5 minutes for the single-session baseline.
Those numbers carry a large caveat, stated plainly by DeepSeek: the tests used synthetic messages and simulated tools, and real coding quality was not measured. The speed advantage describes how the system handles a flood of chatter, not whether its fixes are correct, safe, or worth merging. Teams will have to answer that question themselves, on their own repositories, before trusting it with production code.
Research code, not a product
DeepSeek is careful about what this release is — and is not. The repository calls it a "research prototype for reference only, not an official DeepSeek product," warns that it is not production-ready, and notes that it may still contain bugs. Some designs in the essay, including an external Watchdog, are future directions not implemented in the released code. The team lists open problems familiar to anyone running agents today: Router scaling, long-term memory, and prompt-injection risk when a cloud agent can reach credentials on local machines.
Even so, the release matters for what it signals: coding agents are shifting from models that answer when prompted to systems that notice work and quietly get on with it — and DeepSeek has just published its sketch of that future, free for anyone to inspect and improve.
