DeepSeek’s Harness team published a technical proposal and open-sourced under MIT a research Super-agent prototype for GitHub repositories. A persistent router reads requests and farms work to parallel worker sessions; in preliminary experiments the slowest 5% of tasks finished in about 15.7 minutes versus nearly twice that for a single-session baseline. Maintainers stress that this is research code, not an official product, and that it acts only on explicit requests or team-written rules. By default it does not merge its own code.

What it means: The release shows how Chinese lab teams are pushing toward agents that coordinate many jobs at once instead of waiting for every user prompt. Splitting a router from workers reduces the chance that one heavy task stalls the whole system and makes failure recovery easier to observe. For software teams in Greece, the point is not to copy the prototype blindly but to judge whether a similar pattern can run on internal repositories with clear human review before any merge. An MIT licence lowers the legal barrier, yet security remains open: an agent with access to device credentials or workplace chat can leak secrets unless sandboxes and action logs are in place. Maintainers warn that the public code differs from the internal version, so it should not be treated as production-ready. In the market, the move intensifies competition with Western coding-agent vendors and pushes companies to define who may approve automated pull requests. The sensible path is a pilot on non-critical repositories, mandatory human acceptance, and tight limits on which tools the agent may call. Teams should also budget for router scaling and for collaboration patterns outside a single session, both listed by the authors as open research questions.

Read the original on DeepSeek Harness / South China Morning Post coverage

Related service: Custom software

Related stories