When a workflow requires an agent with conversational capabilities—capable of invoking knowledge bases or plugins—to handle a task, the Agent Creation node dynamically creates such an agent within the workflow by configuring the model, prompt, knowledge base(s), and plugin(s). This agent processes input and returns results, but exists only during the workflow’s execution and is not saved as a standalone agent application.
Usage Example
Create a translation agent using the Agent Creation node to translate Chinese text into English and output the result.
Adding the Node
Method 1: Add via the + Button
- Hover over any node on the canvas (e.g., the Start node); a
+button automatically appears on its right side. - Click the
+button, then select Agent Creation from the pop-up node list. The node will be added to the canvas and automatically connected to the upstream node.
Method 2: Drag from the Node Library
- In the left-hand node library, locate and select Agent Creation.
- Drag the node to a suitable position on the canvas, then drag a connection line from the edge of an upstream node to this node.
Core Configuration
Input
Specifies the content the agent needs to process. Data is passed to the agent by referencing output variables from upstream nodes.
Agent Name
Assigns a unique identifier to the agent, helping distinguish responsibilities among multiple agents in complex workflows.
Model Configuration
Model Selection
Select a large language model (e.g., Qwen series) from the dropdown model selector. If your target model isn’t listed, click More Models, then search for and select it in the modal dialog.
Model Parameters
Click the configuration icon (⚙️) next to the model selector. After enabling the toggle, configure the following parameters:
| Parameter | Default | Description |
|---|---|---|
| Max Response Length | 1024 | Maximum number of tokens in the model’s output. |
| top_p | 0.80 | Controls output diversity; higher values yield more varied results. |
| temperature | 0.70 | Controls output randomness; higher values yield more stochastic outputs. |
| enable_thinking | Enabled | Enables deep reasoning mode, returning step-by-step reasoning traces. |
| thinking_budget | 4000 | Maximum number of tokens allowed for chain-of-thought reasoning. |
| result_format | message | Format of the returned result. |
| enable_search | Disabled | Enables web search capability. |
Prompt
The system prompt is the primary mechanism for controlling agent behavior. Through the prompt, you can define the agent’s domain expertise (e.g., “You are a translation assistant”), specify output format requirements (e.g., “Output only JSON”), enforce behavioral constraints (e.g., “Do not fabricate information”), and more—enabling the same underlying model to exhibit distinct professional capabilities across different scenarios.
The platform provides a Prompt Optimization feature that automatically refines prompts to improve performance.
Knowledge Base
Connect one or more knowledge bases to empower the agent with domain-specific information, enabling accurate, context-aware responses.
| Parameter | Type | Default | Description |
|---|---|---|---|
| Knowledge Base Toggle | Boolean | Disabled | When enabled, the agent can retrieve and reference knowledge base content. The prompt is automatically augmented with knowledge-related instructions and variable references (e.g., ${documents}). |
| Select Knowledge Base | Array | — | Choose up to five knowledge bases to use. |
| Retrieval Chunk Count | Number | 5 | Number of document chunks retrieved per query (range: 1–20). Increasing this adds context but also increases token usage. |
| Show Source Attribution | Boolean | Disabled | When enabled, the agent includes citations indicating which knowledge base entries informed its response. |
Plugins
Plugins extend the agent’s capabilities to perform concrete tasks—such as code execution, web search, or text-to-image generation. You may use both official Bailing plugins and custom plugins.
Node Output
| Variable Name | Type | Description |
|---|---|---|
| result | String | Textual output generated by the agent, containing its processed response to the input. |