DeepSick Operating Prototype
An open-source super-agent architecture that uses a two-layer decoupled Router-Worker design and supports long-term online operation, event-driven behavior, and proactive task execution, aiming to enable continuous autonomous work by AI.
Timeline (last 90 days)
Developer Liangsheng (Liang Sheng) recently open-sourced a super Agent architecture named DeepSick Operating Prototype.
The architecture uses a two-layer decoupled design of a Router layer and a Worker layer: the Router layer handles global scheduling and user interaction, while the Worker layer each maintains independent context and processes specific tasks in parallel.
The Agent no longer relies only on user instructions to start tasks, but listens to external events such as GitHub, group chats, and CI to trigger its working state.
When there is no event worth handling, the Agent stays quiet, consuming no resources and disturbing no one.
The architecture uses tiered cost control: proactively advancing when rules allow, filtering simple events out directly with code, and handing only complex events to the model for judgment, to reduce unnecessary inference overhead.
The real value of the architecture lies in recombining engineering patterns such as Router/Worker layering, event-driven operation, and tiered filtering into the Agent runtime, freeing the Agent from the lifecycle constraints of a single conversation window.
The division of "light global state + heavy local context" is a common pattern in distributed systems, allowing concurrent scaling without linearly raising the context burden at a single point, while ensuring that global consistency is maintained by a single scheduler.
Filtering with code means using deterministic rules (such as event type whitelists, keyword matching, changed file paths, status code judgment) to first perform a layer of screening before entering the model, directly handling or discarding events that clearly need no processing or can be handled programmatically.
All Facts (8)
Filtering with code means using deterministic rules (such as event type whitelists, keyword matching, changed file paths, status code judgment) to first perform a layer of screening before entering the model, directly handling or discarding events that clearly need no processing or can be handled programmatically.
50%UnverifiedThe architecture uses a two-layer decoupled design of a Router layer and a Worker layer: the Router layer handles global scheduling and user interaction, while the Worker layer each maintains independent context and processes specific tasks in parallel.
50%UnverifiedThe Agent no longer relies only on user instructions to start tasks, but listens to external events such as GitHub, group chats, and CI to trigger its working state.
50%UnverifiedWhen there is no event worth handling, the Agent stays quiet, consuming no resources and disturbing no one.
50%UnverifiedThe architecture uses tiered cost control: proactively advancing when rules allow, filtering simple events out directly with code, and handing only complex events to the model for judgment, to reduce unnecessary inference overhead.
50%UnverifiedThe real value of the architecture lies in recombining engineering patterns such as Router/Worker layering, event-driven operation, and tiered filtering into the Agent runtime, freeing the Agent from the lifecycle constraints of a single conversation window.
50%UnverifiedThe division of "light global state + heavy local context" is a common pattern in distributed systems, allowing concurrent scaling without linearly raising the context burden at a single point, while ensuring that global consistency is maintained by a single scheduler.
50%UnverifiedDeveloper Liangsheng (Liang Sheng) recently open-sourced a super Agent architecture named DeepSick Operating Prototype.
50%