DeepSeek describes a Super-agent: one AI dispatches work, several others carry it out

A dispatcher and its workers

As Jasmine News reported, DeepSeek's Harness team published a technical article at the end of October on a "Super-agent" design. It splits the work between two layers: a Router that never stops running, and a set of Workers that do the jobs.

The Router collects incoming information from group chats, GitHub (a code-hosting site) and email, and decides who should handle each item. Each Worker runs in its own isolated session, finishes the task and reports to the Router. The Router is the only part that talks to the user.

Why not just one assistant?

The team's reasoning is that a person can follow only a handful of active sessions at once. Today's mainstream agents treat each session as a separate unit, so someone running several jobs must switch between them and keep each one's context alive by hand.

The team wants an agent that is continuous, efficient, proactive and persistent. In practice, messages, calendar reminders, code-review comments and test results all count as triggers. Simple programmed rules filter them first, and only then does a large language model decide whether to act.

Working ahead, and watching the workers

The system can start work before anyone asks. If a group chat is discussing a bug, it may try to reproduce the fault and look for a fix, and it only reports back once it has confirmed the problem and found a simple repair.

The thinking and session memory live in the cloud, while commands can run on the user's own machine through a small helper program. A separate Watchdog session uses a model to judge whether workers are making real progress or drifting off target.

What the tests do and don't show

In early experiments, the Super-agent finished tasks faster than a single-session setup, including in its slowest typical runs. The team says the tests used synthetic workloads and simplified environments, so they judge scheduling and teamwork, not how well any one task is done.

There is a cost: several sessions mean more cached reading by the model. Removing the direct event pass-through raised total cache reads by more than a third.

Open problems

DeepSeek says an internal Super-agent built this way is already running and helping with bug reproduction, fixes and responses to code review. The team also concedes that connecting cloud agents to local devices raises unresolved risks: prompt injection (hidden instructions that hijack the AI), leaked credentials and permissions crossing devices. It frames the work as applying classic computer-systems ideas, not inventing new theory.