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)
According to a Bilibili uploader's breakdown, the demonstration cases of Super Agent are concentrated in scenarios such as group chats, GitHub, email and CI.
The Router of Super Agent is deliberately designed to be unable to execute tasks itself, in order to avoid a dispatcher bottleneck.
DeepSeek open-sourced a new architecture called Super Agent, which includes two roles: Router (dispatcher) and Worker (executor).
The Router/Worker division of labor in Super Agent provides a scalable skeleton, but because the second type of Worker (mobile apps) is missing, it is currently still out of reach for ordinary users' daily work.
The architectural design of multi-Agent collaboration (such as the Router/Worker division of labor in Super Agent) and the ability to operate real-world interfaces are two independent fronts, and the latter is still in an early exploratory stage.
The capability boundary of Super Agent depends on whether the target system provides programmatic calling interfaces (APIs, command lines, etc.) rather than on the model's capabilities themselves.
Super Agent uses an event-driven input approach, collectively calling group chat messages, code review comments, CI run results, email notifications and the like Triggers.
By assigning unique IDs to Triggers, Threads and Work Sessions and building a reference graph, Super Agent achieves chained recall and supports on-demand retrieval of historical information.
Super Agent introduces a speculative execution mechanism, in which it attempts tasks in the background before results are adopted and, if it fails, silently produces no output.
In context management, Super Agent adopts Inline and Spill strategies: Inline places recently used information directly into the context window, while Spill keeps overly long content outside the window and retains only summaries and IDs.
39 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%UnverifiedDeepSeek open-sourced a new architecture called Super Agent, which includes two roles: Router (dispatcher) and Worker (executor).
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 Router of Super Agent is deliberately designed to be unable to execute tasks itself, in order to avoid a dispatcher bottleneck.
50%UnverifiedThe architectural design of multi-Agent collaboration (such as the Router/Worker division of labor in Super Agent) and the ability to operate real-world interfaces are two independent fronts, and the latter is still in an early exploratory stage.
50%UnverifiedThe Router/Worker division of labor in Super Agent provides a scalable skeleton, but because the second type of Worker (mobile apps) is missing, it is currently still out of reach for ordinary users' daily work.
50%UnverifiedAccording to a Bilibili uploader's breakdown, the demonstration cases of Super Agent are concentrated in scenarios such as group chats, GitHub, email and CI.
50%UnverifiedThe capability boundary of Super Agent depends on whether the target system provides programmatic calling interfaces (APIs, command lines, etc.) rather than on the model's capabilities themselves.
50%UnverifiedIn context management, Super Agent adopts Inline and Spill strategies: Inline places recently used information directly into the context window, while Spill keeps overly long content outside the window and retains only summaries and IDs.
50%UnverifiedBy assigning unique IDs to Triggers, Threads and Work Sessions and building a reference graph, Super Agent achieves chained recall and supports on-demand retrieval of historical information.
50%UnverifiedSuper Agent introduces a speculative execution mechanism, in which it attempts tasks in the background before results are adopted and, if it fails, silently produces no output.
50%UnverifiedSuper Agent uses an event-driven input approach, collectively calling group chat messages, code review comments, CI run results, email notifications and the like Triggers.
50%UnverifiedIn the code, the Router main painting has 14 application tools, but there are no Write, Edit, Blob, or Grab tools, and no Subagent.
50%UnverifiedExperiments show that of about 300 messages per round, only 20 are work requests, and after three filtering stages the vast majority of noise does not enter the model.
50%UnverifiedBefore messages enter the Agent system, programmatic rules first filter the source, then policies filter authorization, and only finally does the model judge whether it is a work request.
50%UnverifiedThe first design principle of Super Agent is "exchange the least human attention for the greatest value".
50%UnverifiedIn the synthetic experiment (each scheme run 5 times), the 95th percentile of task completion time dropped from 29.5 minutes to 15.7 minutes.
50%UnverifiedA Worker reporting Ready does not mean the task is finished; tickets with a PR must wait until the PR is merged or closed to count as finished.
50%UnverifiedThis experiment only measures scheduling and collaboration efficiency, not code quality.
50%