AI Agent
Description
Makes a single AI request, optionally using connected tools, and returns the generated Final Answer. Use it for classification, extraction, summarization, or a bounded task within a workflow. It does not maintain conversation history between executions or stream a conversation.
For an ongoing conversation, use Conversational Agent.
Configuration
Configure a model connection using your provider credentials or an Integration Gateway service. Gateway services consume IB-X Currency.
| Property | Description |
|---|---|
| User Prompt | Required request for this execution. Bind source text or business values from Process Data. |
| System Prompt / Instructions | Behavioral rules, output format, and tool-use guidance for this request. |
| Connection Name | Connection used to access the selected model. |
| Model | Non-realtime chat model supported by the connection. Realtime models are not supported by this activity. |
| Files | Optional collection of file references to include in this request. |
| Temperature | Optional sampling setting, where supported by the model. |
| Max Tokens | Optional positive whole-number output limit. Model-specific upper limits still apply. |
Tools
Connect the required tool activities through Tools and configure their fixed inputs. Expose only the input values the model should supply using the designer's AI-input controls. The model chooses whether to invoke a connected tool from the prompt, instructions, and tool descriptions.
For retrieval, connect Knowledge Base and instruct the agent to search it before answering knowledge-specific questions. A normal execution connection is not a replacement for the Tools connection.
Output
Final Answer is the response text for this request. Its structure depends on the instructions; requesting JSON does not turn it into a typed DataTable. Validate or parse the text before using it as structured input.
A blank prompt or missing usable connection can produce an empty Final Answer. Check that the output contains the information required by subsequent activities. An unsupported realtime model returns an explanatory answer instead of performing the request.
Example: Classify feedback
- Select a model connection and a non-realtime model.
- Set Instructions to: “Classify the feedback as Positive, Neutral, or Negative. Return only the label.”
- Set User Prompt to the feedback text from Process Data.
- Run with representative feedback and inspect Final Answer.
- Route the result using Switch Cases, including a Default branch for an unexpected or empty answer.
For file-based requests, build a file-reference collection and bind Files. See File Reference.
Execution settings
Use Execution Control and On Error for path synchronization and activity failures.