DeepSeek Orb
A desktop agent application built on top of DeepSeek Harness, featuring a floating-orb UI design. It supports a dual-track agent architecture that allows users to control computer operations through natural-language instructions, with a foreground floating-orb agent and a background agent capable of executing tasks in parallel.
Timeline (last 90 days)
DeepSeek Orb is a floating-sphere desktop agent product modified by a Bilibili UP host based on the officially open-sourced DeepSeek Harness, retaining the Harness core loop while adding a floating-sphere UI layer and dual-track scheduling logic.
DeepSeek Orb is structurally divided into two parts: a main window and a floating-sphere window, where the main window is the official Harness itself and the floating-sphere window is a mode for operating the computer.
DeepSeek Orb adopts a dual-track agent architecture consisting of a floating-sphere agent (foreground) and a background agent, with complex long tasks handed off from the foreground to the background for execution to avoid occupying the foreground.
The author mentioned that DeepSeek Orb's background agent is generally paired with models such as Zhipu GLM-4.5 Flash.
In the dual-track architecture, the background agent is suited for pairing with stable, fast, and economical models to execute already-decomposed subtasks, while the foreground agent needs to understand ambiguous natural language and decompose tasks, making it better suited for more capable flagship models.
All Facts (5)
DeepSeek Orb is a floating-sphere desktop agent product modified by a Bilibili UP host based on the officially open-sourced DeepSeek Harness, retaining the Harness core loop while adding a floating-sphere UI layer and dual-track scheduling logic.
50%UnverifiedDeepSeek Orb is structurally divided into two parts: a main window and a floating-sphere window, where the main window is the official Harness itself and the floating-sphere window is a mode for operating the computer.
50%UnverifiedDeepSeek Orb adopts a dual-track agent architecture consisting of a floating-sphere agent (foreground) and a background agent, with complex long tasks handed off from the foreground to the background for execution to avoid occupying the foreground.
50%UnverifiedThe author mentioned that DeepSeek Orb's background agent is generally paired with models such as Zhipu GLM-4.5 Flash.
50%UnverifiedIn the dual-track architecture, the background agent is suited for pairing with stable, fast, and economical models to execute already-decomposed subtasks, while the foreground agent needs to understand ambiguous natural language and decompose tasks, making it better suited for more capable flagship models.
50%