DeepSeek's Super-agent design lets AI sort and start tasks without human prompts
DeepSeek's Harness team published a design in which a coordinating layer handles the juggling of tasks that people now do by hand. An early open-source prototype is on GitHub.
The short version
- The design splits work between a Router, which talks to the user, and Workers, which each execute one task.
- Inputs are events like chat messages, code comments, calendar reminders and test results, not typed instructions.
- Events pass through two filters: fast programmed rules first, then a model decides if action is worthwhile.
- A prototype, dsh-superagent 0.1.0, was uploaded to GitHub under the MIT license.
- The article says a person can juggle only three to five active sessions before mistakes creep in.
Why it matters
Most AI agents today work one conversation at a time, so the human becomes the bottleneck who decides what to run next. DeepSeek is betting that a system can take over the scheduling, filtering and supervising. Whether that frees attention depends on how cheaply you can check the results.
What DeepSeek published
The DeepSeek Harness team put out a technical article on its official WeChat account describing an architecture it calls Super-agent. As China.com reported, the aim is to have AI do the work itself so that human attention is freed up. The same day, an early prototype called dsh-superagent 0.1.0 was uploaded to GitHub under the MIT license, a permissive open-source license.
The problem: people as the bottleneck
The article argues that a person can follow only three to five active sessions at once. Beyond that, context gets mixed up, tasks are missed and priorities go wrong.
Because mainstream agents treat a single session as their basic unit, the number of things they can do in parallel is capped by how many sessions a human can keep in their head. Super-agent tries to hand that mental labor (scheduling, filtering, supervising) to the system.
Router, Workers and tickets
The Router runs continuously. It talks to the user, keeps the overall context and splits work into tickets. Each ticket goes to a Worker, which carries out that one ticket, and its instructions cannot be changed while it runs.
The Router uses a lightweight model to make its judgments, while Workers do the actual execution. A subscription mechanism lets a Worker receive follow-up comments directly, which cuts message overhead. The design also separates compute from storage so tasks can continue after a device goes offline.
Events instead of instructions
Super-agent reacts to things happening around it: group chat messages, code comments, calendar reminders and automated test results. These signals are noisy and only partly structured, so the system filters in two steps. Simple programmed rules discard the obvious noise, then a model decides whether what remains deserves action.
One internal example: when a teammate reports a software problem in a group chat, the system can spot a likely bug without being tagged, then try to reproduce it and look for a fix in the background. A human still makes the final technical call.
What to keep in mind
This is a design description and a version 0.1.0 prototype. The report we saw gives no performance results. The real test is whether checking the system's work costs less effort than doing the work yourself.
What to watch
- Whether DeepSeek publishes results or a demo showing Super-agent working beyond the 0.1.0 prototype.
- Where a person can still approve or stop an action, and how failures are logged.
Try it
Pick one repetitive task, write down what 'done correctly' means, and time how long checking an AI's output takes compared with doing the task yourself.
Sources and links
- ReportingChina.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by China.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.