DocumentationWorkflowsAI and agents
AI Agent node
Calls a model and answers the user prompt. Optionally with tools (Microsoft Graph, HTTP, calculator, …) that the agent picks on its own. Learn how to configure it, map inputs and outputs, run a test and resolve common errors.
8 min readUpdated
What this node does
Calls a model and answers the user prompt. Optionally with tools (Microsoft Graph, HTTP, calculator, …) that the agent picks on its own.
Use AI Agent as a clearly defined step in an anymize workflow.
- Processes one input.
- Provides one output for following steps.
- Can call approved tool nodes autonomously.
Set up the node
- 01Open a workflow and add AI Agent from the node catalog.
- 02Complete the required fields and map values from previous steps.
- 03Run the node with a small test record and inspect its output.
- 04Connect the successful output and activate the workflow after an end-to-end test.
Operations and modes
AI Agent has one focused purpose. Its available settings appear when you select the node.
Settings
| Field | Meaning | Area |
|---|---|---|
| Model | Anymize Waterfall is the EU-hosted default (privacy-first). Claude models for trickier tasks. | Standard |
| User prompt (required) | Configure this value in the node panel. | Standard |
| Output format | Choices: Text (default), JSON (structured) | Standard |
| JSON structure (optional) | Configure this value in the node panel. | Standard |
| Tools | Tools the agent can call. The LLM decides on its own when each one fits. | Standard |
| System message | Prepended to the model. Steers personality + rules. | Standard |
| Maximum reasoning steps | Maximum model rounds (1 to 25). Each round can call several tools. The run fails if the agent does not produce a final answer. | Standard |
| Credit budget | Stops further model and tool calls at this usage (1 to 100000 credits). A model call already in progress can exceed the budget. | Standard |
| Maximum tool calls | Limits each agent step to 1 to 100 tool calls, including multiple calls requested in a single model round. | Standard |
Inputs and outputs
| Input | Meaning | Type |
|---|---|---|
| user prompt | The user message to the agent. Overrides the inline prompt field. | text |
| system prompt | System message / role. Overrides the inline system message. | text |
| Output | Meaning | Type |
|---|---|---|
| output | The agent's answer (final assistant message). | text |
| steps | Configure this value in the node panel. | json |
Example and test run
Research agent with calculator + HTTP tool: begin with a small, recognizable record. Inspect the output and only map fields that are present there.
Test-run example
{
"userPrompt": "Wie viele Einwohner hat Berlin laut Wikipedia, und wie viel ist das mal 0.2?",
"systemPrompt": "Du bist ein präziser Recherche-Assistent. Antworte kurz auf Deutsch."
}Access and security
Before activation, review which data enters this step and what its output contains.
Store keys and credentials in the protected connection manager. Never paste them into normal workflow fields, test data or descriptions.
Troubleshooting
- No output: inspect the latest run and confirm that the previous node returned the expected fields.
- Empty variable: open the previous step output and use a field name from the real test data.