Why DeepSeek Calls It Super-agent: The Design Choices Behind the Prototype
A DeepSeek product manager explains the name, the case for two layers, and the problems the team still needs to solve.
In brief: DeepSeek’s Super-agent puts a coordinating layer above the agent doing the work. In a separate explanation, DeepSeek product manager Cui Tianyi says the design responds to a tension between continuity across tasks and the ability to handle work concurrently. The public implementation is an early research prototype; several ideas in the design remain unfinished.
About this account: This article paraphrases a Chinese-language explanation by Cui Tianyi, who identifies himself as a member of DeepSeek’s product department and a final reviewer of the Super-agent design article. He credits the DeepSeek Harness team with the article’s ideas, design, code and experiments, especially Yujia Qiao, the first author. Cameron Voss and Marcus Bell prepared this independent English account; it is not a direct quotation or an article written by DeepSeek.
Why is it called Super-agent?
The name describes where the extra agent sits. According to Cui Tianyi, “super” means above: a coordinating layer sits over the main agent that carries out work. That contrasts with a sub-agent placed below a main agent. The name is not a claim that the system is inherently superior or possesses some special intelligence.
The distinction is architectural. Instead of treating every request as an isolated conversation, the design explores a higher layer that can keep track of incoming work while execution happens in separate tasks.
Want to try DeepSeek now? Explore our DeepSeek subscription.
Is the two-layer idea a new theory?
No. Cui Tianyi says the team did not present it as a new theory, and that other teams had already experimented with similar approaches. In his account, the difficult part is choosing a direction that continues to make sense over time and optimizes the system as a whole, rather than merely trying every possible arrangement of agents.
The released dsh-superagent-prototype is an early exploration. The public repository describes it as research code, not an official DeepSeek product or production-ready software. It also says the published prototype differs from the internal version and the design article; some ideas discussed in the article are directions for future work.
Why add a layer above the working agent?
One motivation is long-term memory. Cui Tianyi says there is still no agreed solution for giving agents continuity across long-running work. At the time of his explanation, DeepSeek Harness did not have built-in long-term memory across sessions; he suggested that might change, without announcing a release.
His framing is a tradeoff: an agent needs continuity across tasks, but it also needs to handle enough work. A single ongoing session may preserve context, while separate work sessions can make concurrent execution easier. The two-layer design is an attempt to reason about that tension, not proof that the team has found the best implementation.
| Design concern | What the explanation suggests |
|---|---|
| Continuity | Keep track of work across tasks instead of relying solely on one short-lived interaction. |
| Throughput | Let execution proceed across tasks while a higher layer handles coordination. |
| Open question | How to preserve useful context without letting coordination or memory become a bottleneck. |
The last row is an editorial interpretation of the tradeoff, not a measured result. The screenshots do not provide a benchmark or establish that this architecture resolves the memory problem.
What remains unfinished?
Cui Tianyi says the team simplified both the engineering and the scenarios to keep its internal system manageable. Moving to broader real-world use would require answers to several open questions raised in section 8 of the design article:
- How to extend the two-layer architecture.
- How people and agents collaborate beyond a single session.
- Where context boundaries sit, and how agents communicate with one another (A2A).
- How to address security and privacy.
He also notes that separating computation, state and environment would take substantial additional work in DeepSeek Harness. The prototype has been used in a relatively small, trusted internal setting. Real deployments introduce more varied conditions and unresolved problems.
What should teams take away?
Treat the design as a way to examine continuity, coordination and workload, rather than as a finished blueprint. Before applying it to a broader setting, identify who can authorize work, what context each task receives, how execution is isolated, and what happens when a task fails. Those are evaluation questions, not claims that the prototype has solved them.
For a closer look at the public implementation, read our SuperAgent setup and cost guide and SuperAgent explainer.