DSH Super Agent
An open-source multi-agent collaboration framework from DeepSeek that adopts a two-tier Router-Worker architecture, emphasizing the scheduling and restraint of human attention, and includes mechanisms such as subscription passthrough, speculative execution, and Watchdog.
Timeline (last 90 days)
DSH Super Agent's architecture is divided into two layers: the upper Router layer receives messages, decomposes tasks and maintains global context; each lower-layer Worker has its own independent working directory per task, executing in parallel without interfering with each other.
DSH Super Agent uses a "subscription pass-through" mechanism: after a Worker takes over a PR, it directly subscribes to that PR's comment stream, and subsequent comments are pushed directly to the Worker without Router forwarding.
Cancelling the subscription pass-through channel increases the Router's number of decisions by 30% and the tokens for reading the cache by 37%.
DSH Super Agent has two filters at the message entry: programmatic rules first screen out irrelevant messages, then the model decides whether to act; and the AI can ask only one specific question at a time, in the form of a structured form.
DSH Super Agent separates compute, state and execution environment: inference runs in the cloud, session state is stored locally separately, and recovery after shutdown is supported.
The first line of DSH Super Agent's readme states: this is a research prototype, not an official DeepSeek product, it may have bugs, the engineering style is inconsistent, and you need to modify it yourself for production.
DSH Super Agent currently targets only a single GitHub repository, with optional Feishu integration.
Some of the design prototypes mentioned in the DSH Super Agent paper have not yet been implemented.
All Facts (8)
DSH Super Agent's architecture is divided into two layers: the upper Router layer receives messages, decomposes tasks and maintains global context; each lower-layer Worker has its own independent working directory per task, executing in parallel without interfering with each other.
50%UnverifiedDSH Super Agent uses a "subscription pass-through" mechanism: after a Worker takes over a PR, it directly subscribes to that PR's comment stream, and subsequent comments are pushed directly to the Worker without Router forwarding.
50%UnverifiedCancelling the subscription pass-through channel increases the Router's number of decisions by 30% and the tokens for reading the cache by 37%.
50%UnverifiedDSH Super Agent has two filters at the message entry: programmatic rules first screen out irrelevant messages, then the model decides whether to act; and the AI can ask only one specific question at a time, in the form of a structured form.
50%UnverifiedDSH Super Agent separates compute, state and execution environment: inference runs in the cloud, session state is stored locally separately, and recovery after shutdown is supported.
50%UnverifiedThe first line of DSH Super Agent's readme states: this is a research prototype, not an official DeepSeek product, it may have bugs, the engineering style is inconsistent, and you need to modify it yourself for production.
50%UnverifiedDSH Super Agent currently targets only a single GitHub repository, with optional Feishu integration.
50%UnverifiedSome of the design prototypes mentioned in the DSH Super Agent paper have not yet been implemented.
50%