# DeepSeek publishes design and open repo for its Super-agent harness prototype > DeepSeek released a design article and a prototype repo for a long-lived router agent that hands work to isolated workers triggered by GitHub events. - Status: official (From the company, authors or authority themselves, with a link to their own page.) - Topic: Open Source - Published: 2026-11-03T05:35:54+00:00 - Updated: 2026-11-03T05:35:54+00:00 - Web page: https://ai-news.ikangai.com/story/deepseek-publishes-design-and-open-repo-for-its-super-agent-harness-prototype/ - Seen in: Latent Space DeepSeek has published a long design article and a prototype repository, dsh-superagent-prototype, that describes what it calls a Super-agent. It is one long-lived router session that talks to humans and assigns work, with isolated worker sessions running tasks in parallel and triggered by GitHub events. The article is a machine translation of a Chinese original, and an official English version followed. It is explicit about its scope: it discusses architectural ideas and does not introduce features that already exist in DeepSeek Harness. The design rests on a split between two layers. The router never executes tasks itself. It only talks to people and dispatches work, and the article says it can run without a Bash tool at all. Each worker has an isolated context, a full set of execution tools and a single task. Workers never talk to humans directly; they report status back to the router, which handles every reply. The router is one long-lived session so that context stays continuous across channels. Batching is why it is unlikely to become a throughput bottleneck: events that arrive while the router is mid-inference queue up and merge into the next model call. Because the router's calls are mostly long input and short output, the article argues that a backlog costs little extra per event, since merged inputs are processed together during prefill. The article also says a lighter model is usually enough for the router, while workers need the most capable models. Events, which the article calls Triggers, include group chat messages, emails, calendar reminders, code review comments and CI results. They pass through two filters before reaching a model. Deterministic code first checks the source, the group and allowlists, and drops most noise. Remaining candidates go to an LLM that decides, based on prompt-written policies, whether to act. Follow-up events in the same thread, such as a GitHub pull request or an IM topic, go straight to the worker subscribed to that thread, so they skip the router. The article presents this as a fast path that keeps the single global node from becoming a bottleneck. The main quantitative claim comes from DeepSeek's own test: p95 completion time fell from 29.5 to 15.7 minutes compared with a single session using subagents. The article also offers an anecdote from DeepSeek's internal Super-agent instance. It says the instance continued one development task across IM and GitHub and, even without being @-mentioned in a public group chat, found a root cause on its own and opened a pull request to fix it. That is a single illustrative case from the authors, not a measured result. Zhihu Frontier published an explanation by 崔添翼 (@tianyi), whom it identifies as a member of DeepSeek's product department and one of the article's final reviewers. He says 'Super' means a layer above the working agent, not superior intelligence, and frames the core trade-off as continuity across tasks against throughput. The article makes the same point in its own terms. Sending every event to one session gives the most continuous context but caps throughput, because a ReAct loop runs one step at a time and task details interleave. Giving every event a new session maximizes concurrency but leaves related results disconnected and confuses users who receive several separate replies. The article lists its own limits. It has no built-in cross-session long-term memory yet, and separating state, compute and environment is mostly unimplemented. The prototype has been validated only in a small, trusted internal environment. Security, privacy and agent-to-agent communication remain open questions. The article also says that sharding the router would bring back the continuity-throughput trade-off, which will remain until the global memory problem is solved. On prior work, the author acknowledges that other teams explored similar architectures months earlier. Coverage and reaction have been mixed. Bloomberg reported on November 1 that DeepSeek had released a prototype of a more autonomous coding agent as it seeks to challenge Anthropic's Claude Code. Teortaxes noted that the repository was dropped without any English announcement and was still covered by Bloomberg. The same commentator called the write-up 'something of a hot mess' and argued that DeepSeek builds internal crutches around model weaknesses such as hallucination rather than making the system generally robust. On the other side, @badlogicgames said the design matches where Earendil is heading, and @omarsar0 called it a 'masterclass' in coordination-execution harness design. According to @tianyi, DeepSeek also says it aims to support pi and other major harnesses with each model generation. ## Why it matters Companies building multi-agent coding workflows get a concrete, open reference for routing work across parallel workers, but the stated gaps in memory and security mean it is not yet production ready. ## What you can do Teams evaluating agent orchestration should review the stated limits, especially memory and security, before adopting the pattern. ## Original source - [DeepSeek Harness on X: "Exploring Super-agent Design: Towards a More Autonomous Agent System" / X](https://x.com/DeepSeekHarness/status/2116857257245987236) - [Zhihu Frontier on X: "Read the full answer by 崔添翼 on Zhihu: https://t.co/dGNv31SlLQ Super-agent prototype: https://t.co/jQPVXN4LIy" / X](https://x.com/ZhihuFrontier/status/2117074614434316705) ## Further reporting - [Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) on X: "little Whale has come so far not long ago their models received no interest now a half-baked prototype from the har](https://x.com/teortaxesTex/status/2116856872003375236) --- AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.