What happened
On October 31, 2026, the deepseek-ai organization published dsh-superagent-prototype on GitHub under the MIT license. The README labels it a research prototype for reference only, not an official DeepSeek product, adapted from internal day-to-day work and not identical to the system in Exploring Super-agent Design: Towards a More Autonomous Agent System. It is packaged as a bundle for DeepSeek Harness. It listens to GitHub and optionally Feishu. When someone allowed asks for work, it opens a work item, works in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person. In the architecture note DeepSeek Harness posted with a machine translation, DeepSeek reports a synthetic end-to-end schedule where Super-agent finished in 12.4 minutes at p50 and 15.7 at p95, versus slower single-session baselines in that same setup, and says that without pass-through the Router made about 30% more decisions while cache-read tokens rose from 23.2M to 31.8M, about 37% more. DeepSeek says those experiments use synthetic workloads and are indicative only.
Why it matters
DeepSeek already sells models. This repository and note show how it wants a coding agent to sit around those models: a Router that dispatches, Workers with isolated contexts, and rules for when to spend a person's attention. The 12.4-minute and 37% figures are DeepSeek's synthetic schedule results, not a public coding benchmark. The README still says treat the code as research with possible bugs.
DeepSeek Harness itself is described on deepseek.com as a public-preview open-source desktop and web harness with composable plugins.
DeepSeek says Super-agent without pass-through finished tasks in about the same time but burned more Router decisions and cache-read tokens.
The README points at numbered design docs for philosophy, authorization, execution, prompts, panel, and deployment inside the repo.