Free Handoff is a model-driven task transfer method. Its basic feature can cover most task scenes. When creating a Multi-Agent Mode application, the collaboration method is selected by default as Free Handoff, and a primary Agent is created by default.
The main advantage of free transfer is simple configuration. By setting the transfer relationship, the model can autonomously judge when to transfer, offering extremely high freedom. The primary defect is instability caused by high freedom, requiring relatively high configuration requirements for Agent prompt content.
In Free Handoff collaboration mode, it supports creating one Agent or more. Multiple agents complete tasks through transfer collaboration.
Setting Method
Step One: Configure Agent
1. Configuration Name
When switching to Multi-Agent Mode, the system automatically creates a Main Agent with the same name as your application. Users can edit this name by clicking the edit icon.
Note:
For single-Agent apps, a neutral name is fine. For Multi-Agent apps, both the agent name and description are key information for task transfer. Use simple, meaningful names to describe agent features, such as Travel Planning Assistant, Web Page Analytics Agent, or Copywriting Optimization Agent.
2. Configuration Model
The model in the Agent is used for task planning and tool selection, supporting model selection as needed.
The maximum number of reasoning rounds represents the maximum number of times the model can "think + make tool calls". It is advisable to set it reasonably based on task complexity to balance performance and resource consumption.
Decrease: Reduce token consumption and reasoning time, suitable for simple tasks.
Increase: Supports more complex call scenarios but consumes more tokens.
3. Transfer Configuration Description
The Agent description briefly explains this Agent's feature and use cases, helping other Agents judge when to transfer tasks to it.
Note:
If configuring a single Agent usage scenario, you can simply fill in the transfer description. But in multi-Agent collaboration scenarios, clear descriptions are critical for correct transfer.
4. Configuring Prompt Content
Support adding the task objective, task flow, and limitation notes of the current Agent in the prompt content.
You can refer to the Template and fill in according to actual needs, or use the AI one-click optimization feature to quickly add related prompts and improve content.
5. Configuring Plugin
A tool is an API that fulfills a requirement by calling it. A plugin is a collection of tools, where a plugin includes several tools (essentially a tool group). In an Agent, the tools selected are those included in the plugin. Model invocation of tools involves registering the tool's name and description in the system prompt of the large model, enabling the model to learn which tools are available and their specific usage scenarios. When a user enters a question, the large model selects one or more suitable tools to resolve the issue based on its understanding and breakdown of the question, as well as the tool descriptions.
Therefore, for an Agent, it is necessary to select the required tools and provide easy-to-understand tool descriptions (pay attention when using custom plugins), which is critical for successful tool calls.
Tool support for configuring parameters. After adding a plugin, click Tool Settings.
In Tool Settings, under Input Parameters, you can enter or refer to parameter default values, while also supporting switching whether the model is visible for a single parameter. If set to visible, the model collects this parameter in conversation; if set to invisible, the tool is called with the imported or input parameter.
In addition to supporting input parameter values, you can also dynamically pass in parameters via referencing API parameter values through the API.
In Advanced Configuration, you can configure API parameters. Click Create to set the parameter name, parameter description, and parameter type for the configuration of API. The configured parameters can be referenced in tool parameters.
When calling the Tencent Cloud Intelligent Agent development platform API, pass variables through the custom_variables field (for details on this parameter usage, see Dialog API Documentation (HTTP SSE), Dialog API Documentation (WebSocket)), then reference this variable in tool parameters to execute subsequent business logic.
Note:
For example, in the "Permission Query" Agent, when calling the Tencent Cloud Intelligent Agent development platform API, you can place the userID (assuming this field represents the employee ID) into the custom_variables field and pass it to the Agent via a custom plugin to achieve different permission scopes for different userIDs. Configure the userID field in the advanced setting's API parameters, ensuring the field name and data type remain consistent with those passed in the API. During plug-in configuration, you can directly reference this variable. At runtime, the system will automatically parse and use the userID field in custom_variables to call the corresponding plugin.
The debugging parameters in the top-right corner of the debug dialog box support direct configuration of custom API parameter values, making it easy to perform Agent feature debugging.
During debugging, modify the default value of the corresponding parameters to refer in the plug-in's tool settings, select the custom API parameter, and set model not visible.
Step Two: Add an Agent
The application supports adding multiple Agents to respond to dialogue. If only a single Agent scenario is needed, no need to add extra Agents.
The range of Agents that can be added includes ALL Agents in other "Multi-Agent mode" applications.
Click Add to support editing the Agent before adding it to the application.
Step 3: Transfer Relationship Configuration
When adding two or more Agents in the application, the system supports configuring transfer relationships between multiple Agents. Please note, by default, these Agents will not take effect immediately after creation. The Main Agent can accurately allocate dialogues to specific sub-Agents based on "Agent description" only after establishing transfer relationships with them. If a sub-Agent fails to establish an association with the Main Agent through transfer relationships (which can be single-layer or multi-layer), it will be deemed as in a free status and cannot be activated for use.
The recommended configuration is a centralized architecture where the Main Agent transfers tasks to multiple sub-Agents, while each sub-Agent also transfers back to the Main Agent.
