"Multi-Agent Mode" is one of the application categories supported by the intelligent agent development platform. Applications in this mode feature autonomous execution path planning by large models and flexible tool invocation, with the core characteristic of shifting more dialogue initiative to the model to fully leverage its proactiveness. This mode is suitable for scenarios requiring flexible response, multiple tool calls, and multi-agent collaboration.
In Multi-Agent Mode, it supports creating single-Agent applications, allows adding multiple Agents, and supports switching between different multi-Agent collaboration methods to satisfy multi-scenario requirements.
Key Concepts
Concept
Definition
Name
Simple and clear name description for Agent features, such as: travel planning assistant, web page analysis Agent, copywriting optimization Agent.
Model
A large model responsible for thinking, task planning, and tool selection. Supports different model options.
Transfer description
Briefly explain the Agent's feature and use cases to help other Agents judge when to transfer tasks to it.
Prompt
Use prompts to define the Agent's task flow execution and response method. Unlike the "Agent description," the "prompt content" helps the model understand and execute the current Agent's detailed work logic.
Plugins
Plugins represent a collection of tools, supporting adding tools from the plugin center (including tools in MCP Server) to the current Agent.
Transfer relationship
Which other Agents can the Agent transfer to?
Collaboration method
The transfer methods for Agents include freely transfer, workflow collaboration, and Plan & Execute.
Core Workflow
Multi-Agent Mode includes three types of core tasks: thinking and planning, proactive tool selection and calling, and proactive error correction and reflection.
Thinking and planning: Formulate a planning approach to attain the overall task and deconstruct the complex task into segmented subtasks.
Proactive tool selection and calling: Based on the deconstructed subtasks and tool description, select one or more suitable tools to resolve the issue.
Proactive error correction and reflection: The model autonomously improves and optimizes past decision-making and corrects behavior.
The three types of work may execute cyclically multiple times and eventually output the reply answer.
Core Process Example
A user enters the question "Shenzhen weather for the next week" as an example. The Agent will respond as follows.
1. [Thinking] Step one requires querying the weather information in Shenzhen. And select the GetWeatherInfo tool to query the weather info.
2. [Tool call] Query the weather info for the next week via GetWeatherInfo.
3. [Thinking] The weather results have been obtained. Think about the next action, automatically generate code, and draw the temperature trend chart through CodeInterpreter.
In the reply example for the above issues, the intermediate model plan and tool call process result is "thinking process," and the final output is the answer.
Collaboration Method Selection
The system supports selecting multiple collaboration methods for Agent transfer. For the application creation process of different collaboration methods, see the corresponding documentation.
Model-driven task transfer method with user-friendly operation, but transfer stability depends on the Agent name and clarity level of the transfer description.
By orchestrating Agent nodes with fixed processes, task execution is stable and controllable.
Fixed workflow tasks requiring stable execution
When creating a Multi-Agent application, the default Agent collaboration method is Free handoff. At this point, the transfer process will be autonomously driven by the model. When switching to Workflow Orchestration, the transfer sequence is fixed, and task execution is controllable.
FAQs
Switching From Standard Mode to Multi-Agent Mode, Will the Edited Application Configuration Content Be Retained
App settings are independent across different modes and are not inherited when switching modes, including model selection and prompt content. However, the knowledge base scope and workflow scope remain synchronized between modes.
When switching back from Multi-Agent mode to standard mode again, you can modify the already edited prompts and configurations in the original standard mode.
Difference between Multi-Agent Mode and the Other Two Modes
Multi-Agent Mode differs from standard mode in that execution does not follow a fixed standard process. Instead, it depends on the large model to plan tasks, deconstruct them based on user-submitted input, and proactively call tools. The large model must understand the name and description of tools in plug-ins to choose appropriate plug-ins.
Multi-Agent Mode differs from single workflow mode by not executing a single specific workflow. For deterministic task scenes, we recommend choosing single workflow mode. The current version of Multi-Agent Mode does not support workflows. Stay tuned for future upgrades.
Why Does Multi-Agent Mode Not Support Workflow in Free Transfer Collaboration Method
Multi-Agent Mode supports workflows, but requires consideration for compatibility with multi-round interactions and Agent thinking patterns. Currently, workflows are only supported when workflow orchestration is selected as the collaboration method in Multi-Agent Mode. Meanwhile, we are exploring some interesting features to enable workflow support under free transfer collaboration in Multi-Agent Mode, achieving flexible workflow redirection and undertaking. If interested in this part, please contact operations for consultation.
Is Multi-Agent Mode Only Supports Multi-Agent Configuration
Multi-Agent Mode supports not only multi-Agent collaboration scenarios but also single-Agent scenarios. In single-Agent scenarios, no additional configuration is required, and the system-created Agent can be used directly. To implement multi-Agent collaboration, multiple Agents can be added with transfer relationships configured, thereby facilitating the construction of a multi-Agent system.
