Subagent System

Overview

A subagent is the task decomposition and delegation mechanism of the workspace. The main Agent can break down a complex parallel task into subtasks, create dedicated subagents to execute them independently, and after the subagents finish, the results are returned to the main Agent for further processing.

The system supports three types of subagents, with a maximum nesting depth of 3 levels.

Three Types of Subagents

Automatically Created Subagents (create_local_sub_agent)

When the Agent determines during reasoning that the current task needs to be decomposed, it will autonomously call the create_local_sub_agent tool to create a temporary subagent. Users can also request the Agent to create a subagent through conversation prompts, for example:

"Please create a subagent to search for relevant materials, then create another subagent to write the document"

Creation Parameters:

Execution Flow:

1. The main Agent decides to create a subagent and assigns the task
2. The system starts an independent Agent Loop (default maximum of 50 rounds)
3. The subagent has its own tool loop, and its execution process is visible in the parent conversation window
4. After the subagent finishes, it returns the final result to the main Agent
5. The main Agent continues subsequent work based on the result

Pre-registered Dispatch Subagents (dispatch_sub_agent)

Users can pre-register subagents with specific configurations in the APP settings, and the Agent invokes them via the dispatch_sub_agent tool when needed.

The descriptions of pre-registered subagents are automatically injected into the main Agent's system prompt, so when the LLM sees a matching task, it automatically selects the appropriate subagent.

Configuration Items:

For the management interface, see Subagent Registration and Management.

Cross-node Dispatch (dispatch_multi_node_agent)

When a task requires resources or capabilities of a remote device, the Agent can dispatch the task to another node for execution via the dispatch_multi_node_agent tool.

Workflow:

1. The main Agent views the list of online nodes (injected into the system prompt)
2. Selects the target node based on the node capability declarations
3. Sends node.invoke.request via the Gateway WebSocket
4. The remote node starts an independent Agent Loop to execute the task
5. Returns the result via node.invoke.result upon completion

Scenario Example:

• The Web client on Computer A initiates a conversation
• The task requires accessing local files on Computer B
• The Agent dispatches the file operation subtask to Computer B's APP node via the Gateway
• After Node B executes it and returns the result, the conversation continues to be displayed on Computer A

Nesting Depth Control

Subagents support nested creation, with a maximum depth of 3 levels:

                      
                      深度 0: 主 Agent（用户对话）
  └── 深度 1: 子代理 A
        └── 深度 2: 子子代理 A-1
              └── 深度 3: 最深层代理（不可再委派）

                    

The greater the depth, the stronger the autonomy, but Token consumption and execution time increase accordingly.

Tool Inheritance Mechanism

UI Display

Subagents are displayed with special markers in the Work conversation window:

Users can clearly see which operations are performed by the main Agent and which are performed by subagents.

How to Create a Subagent Through Conversation

You can request the Agent to create a subagent using natural language in the conversation. Here are some example prompts:

Interpreting the Interface During Subagent Execution

Cost Impact

The token consumption of subagents is calculated independently:

Recommendation: Subagents are suitable for complex tasks that genuinely need to be decomposed. For simple tasks, just let the main Agent handle them directly—there is no need to create subagents.

When to Use / Not Use Subagents

It is not recommended to assign different roles to Subagents for a division of labor and collaboration; in the AI era, collaboration actually reduces efficiency and quality. It is more suitable for scenarios of parallel task execution to save time and improve efficiency.

Related Documentation

• Subagent Registration and Management — Pre-registered subagent configuration
• Multi-node Architecture — The underlying mechanism of cross-node dispatch
• Agent Loop Engine — The Agent Loop used by subagents