DeepSeek Opens Up Its Super-agent Blueprint: A Router-Worker Coding Agent That Watches GitHub and Finishes Tasks in a 12.4-Minute Median

DeepSeek has published the design philosophy behind its next generation of agents, and for once the artifact is not a model weight but an architecture. The team released an article titled "Exploring Super-agent Design: Towards a More Autonomous Agent System" together with a working prototype, dsh-superagent-prototype, on GitHub under the MIT license. The repository is explicit about its status: it is a research prototype for reference only, not an official DeepSeek product, adapted for publication and not identical to the internal version — and much of its engineering, the README notes, was done by coding agents.

The problem it targets will be familiar to anyone juggling AI sessions. Real software work sprawls across platforms: a technical decision gets made in a group chat, the code lands in a GitHub pull request, and the review comments arrive somewhere else. Traditional coding agents treat each of those as a separate session, and the human has to remember where everything stands. DeepSeek's observation is that a person can realistically track only three to five active sessions, so as models get better at single tasks, multi-task throughput is increasingly capped by human attention.

The architecture is two-tier. A Router at the top receives messages, assigns tasks and maintains the global context; beneath it, Workers run the actual agent sessions, each in its own directory, so multiple tasks advance in parallel without blocking new interactions. A subscription and direct-pass mechanism keeps old tasks alive: a Worker that takes over a GitHub pull request subscribes to later comments and handles them directly, without routing every message back through the Router. Incoming events pass a two-stage filter — program rules discard the noise first, and a large language model judges only the remainder — which DeepSeek says keeps token spend under control.

The more provocative ideas sit in the proactive layer. Speculative execution lets the agent start working before it is asked: if someone mentions a bug in chat, the Super-agent may reproduce it and hunt for a fix in the background, surfacing the result only if it finds something clean and staying quiet otherwise. An asynchronous ReAct loop lets long tasks run while the agent keeps answering new instructions; DeepSeek says it is implemented in its Harness but not enabled by default. Compute, session state and execution environment are kept separate so a laptop going offline does not kill cloud-side work, and a dedicated Watchdog agent monitors for crashes, stuck loops and goal drift, stepping in to repair when needed.

The numbers are the team's own and should be read as vendor-reported. In a simulated collaboration environment with roughly 300 messages and 20 work requests, each configuration tested five times, the Super-agent completed tasks at a median of 12.4 minutes and a 95th percentile of 15.7 minutes, while a traditional single-session setup with subagents hit 29.5 minutes at the 95th percentile under neutral prompts. The routing design also has a measurable cost: without the direct-pass mechanism, Router decisions rose about 30 percent and cached-token reads about 37 percent.

The prototype itself is scoped tightly: a proactive coding agent for a single GitHub repository, listening to GitHub and optionally Feishu. When an authorized person asks for work, it opens a work item, works in its own session and directory, and answers back in the original conversation, for example with a pull request. A web panel shows what still needs a human, and the repository's philosophy document draws a hard line: the agent acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act or can use a result — because an agent that can act for a whole team can also waste a whole team's attention.

It is a notable inversion of the usual release cycle. Labs typically ship the model and let an ecosystem of harnesses grow around it; DeepSeek, which already maintains the DeepSeek Harness, is publishing the operating manual for always-on, multi-session agents and letting anyone run it. Whether Super-agent-style autonomy survives contact with real teams is the open question — but the blueprint, bugs and all, is now public.