Super Agent
An open-source team-level AI programming system from DeepSeek that runs on DeepSeek Harness, receives tasks via GitHub and Feishu, and has AI agents perform code modifications and return the results.
Timeline (last 90 days)
The four core characteristics of the Super Agent architecture are continuous, efficient, proactive, and persistent.
The Super Agent article is still at the architecture exploration stage, and the specific implementation details and degree of openness remain to be observed later.
Super Agent takes roughly only half the time of the single-session scheme with SubAgent tools.
In measured tests, Super Agent's P50 completion time is 12.4 minutes and its P95 completion time is 15.7 minutes.
Super Agent uses a Thread subscription mechanism: the Worker Agent subscribes to topics directly, and subsequent replies go straight to the Worker rather than being dispatched through the Router.
The Super Agent architecture uses a two-tier structure of Router Agent and Worker Agent, with the Router Agent serving as the user entry point responsible for coordination and the Worker Agent responsible for executing tasks.
Super Agent identifies environmental information that needs handling through two layers of filtering—programmatic filtering (machine learning or code rules) and model judgment—to support proactive execution capability.
Super Agent's memory design includes two mechanisms: Inline (reading task summaries of other recent related Sessions) and Spill (keeping long content outside the conversation window and expanding it when needed).
Super Agent's chained recall mechanism means that the Router locates the Worker's specific execution details when needed and cites them back.
Super Agent's "one-step advance" means that when the Agent receives a new request, it does not interrupt the current flow but processes it in parallel.
18 more timeline events
All Facts (20)
The cost of the Super Agent architecture is that more sessions mean more Token consumption.
60%UnverifiedThe Super Agent architecture uses a two-tier structure of Router Agent and Worker Agent, with the Router Agent serving as the user entry point responsible for coordination and the Worker Agent responsible for executing tasks.
60%UnverifiedThe DeepSeek Harness team has published a SuperAgent design article and GitHub repository, released under the MIT license and explicitly labeled as a research prototype.
60%UnverifiedThe four core characteristics of the Super Agent architecture are continuous, efficient, proactive, and persistent.
50%UnverifiedIn measured tests, Super Agent's P50 completion time is 12.4 minutes and its P95 completion time is 15.7 minutes.
50%UnverifiedThe Super Agent article is still at the architecture exploration stage, and the specific implementation details and degree of openness remain to be observed later.
50%UnverifiedSuper Agent takes roughly only half the time of the single-session scheme with SubAgent tools.
50%UnverifiedSuper Agent uses a Thread subscription mechanism: the Worker Agent subscribes to topics directly, and subsequent replies go straight to the Worker rather than being dispatched through the Router.
50%UnverifiedSuper Agent identifies environmental information that needs handling through two layers of filtering—programmatic filtering (machine learning or code rules) and model judgment—to support proactive execution capability.
50%UnverifiedSuper Agent's memory design includes two mechanisms: Inline (reading task summaries of other recent related Sessions) and Spill (keeping long content outside the conversation window and expanding it when needed).
50%UnverifiedSuper Agent's chained recall mechanism means that the Router locates the Worker's specific execution details when needed and cites them back.
50%UnverifiedSuper Agent's "one-step advance" means that when the Agent receives a new request, it does not interrupt the current flow but processes it in parallel.
50%UnverifiedSuper Agent adopts a cloud-based continuous coordination design, with a cloud computer built into the Agent, so shutting down the local computer does not affect progress in the cloud.
50%UnverifiedSuper Agent sets up an independent Agent Watchdog outside the main system that runs continuously to check for issues such as task progress deviating from the goal and to fix them.
50%UnverifiedAccording to the Bilibili creator's account, Deepseek Harness's measured tests of the Super Agent architecture show that in 95% of cases Super Agent took far less time than the traditional single-Session Agent architecture.
50%UnverifiedAccording to the Bilibili creator's account, Super Agent takes less than 16 minutes on average to complete a task, while the traditional architecture takes close to 30 minutes.
50%UnverifiedAfter Super Agent is connected to a GitHub code repository, teams can assign it work through the discussion area, and it organizes AI to handle it and returns code changes or processing results.
50%UnverifiedSuper Agent gives each task its own session and working directory, with context isolated between different tasks, and subsequent messages of the same task reuse the same session.
50%UnverifiedSuper Agent's internal structure can be divided into three parts: the entry layer receives GitHub and Feishu messages, the scheduling layer determines whether a task is new or a continuation of an original task, and the execution layer operates on code and the environment.
50%UnverifiedSuper Agent has a web panel where task status can be viewed, with priority given to displaying matters that require human handling.
50%