Super Agent01 / 10
Super agent DeepSeek

DeepSeek Super Agent Architecture Explained

A Router-Worker design for a coding agent that can hold more than one task. The public code is a research prototype, not an official DeepSeek product.

Name02 / 10
What the name refers to

An essay, and a narrower repository

The essay

Exploring Super-agent Design describes a continuous agent: one voice, several tasks, events from chat and GitHub.

The repo

deepseek-ai/dsh-superagent-prototype is a TypeScript bundle for DeepSeek Harness. Its banner says research code, not production software.

Not identical

The repo says it was adapted for publication, so it is not the internal system and not every idea in the essay.

Problem03 / 10
The bottleneck

The session became the product

A single chat coding agent stores memory and does the work in one loop. Unrelated tasks contaminate that loop. A fresh session for every event loses the shared picture, and several sessions may answer the same person. The DeepSeek Harness team says its experience is that a person can actively follow about three to five sessions. That is their observation, not a published study.

Split04 / 10
The split

Continuity above, execution below

Router

Long-lived. It keeps global context, talks outward, and dispatches work. The essay says it can operate without doing the task itself.

Worker

One assignment, an isolated context, and the tools to execute. Workers report back so the user hears one voice.

Not a team lead

A team lead decomposes one coupled task. The router decides that a task exists and leaves the decomposition inside the worker.

Path05 / 10
How a task starts

Triggers become authorized work

Trigger

GitHub, and optionally Feishu, produce events. The essay also discusses mail and other channels as direction.

Decision

Something has to decide the event asks for new work, not just noise.

Ticket

Accepted work is a ticket with a reason it may proceed. Each ticket gets its own session and directory.

Return

Later comments, such as on a pull request, can reach the same worker without a new mention.

Engine06 / 10
The engine

State changes in one place

The prototype’s architecture note says Engine is the only writer of Triggers, Threads, Decisions, Tickets, and Assignments. A command changes those records in one transaction. External work is queued inside the transaction and performed after commit. That is the difference between a chat log and a system of record.

Compare07 / 10
The difference

Who schedules the work

One chat agent

You are the scheduler. Each window is a session you must remember.

Router-Worker

The router keeps the relationship across tasks. Workers run apart so one slow command does not freeze the conversation.

Still a person

The prototype acts when asked, or when a rule the team wrote allows it. Proactivity without a policy is just noise.

Prototype08 / 10
The public code

One repository, not the whole essay

Read09 / 10
Takeaways

Use it as a reference design

Read the banner first

Research prototype. Not an official product. Not production-ready.

Separate the layers

Essay for the idea. Repository for what you can run. Engine for how state is saved.

Do not drop it in

A coding agent that can open pull requests needs an authorization model you have actually read.

Sources10 / 10
Sources

Primary notes, then commentary

Public sources, read in November 2026. Projects move. Claims marked as project statements were not remeasured for this deck.