AI Agent

The AI Agent node executes LLM calls with optional tool use. Without tools, it makes a single LLM call. With tools enabled, it becomes an autonomous agent that can reason, use tools, and iterate.

Configuration

Model Selection

Select a provider and model:

Provider	Popular Models
Anthropic	claude-sonnet-4-5, claude-opus-4-5, claude-haiku-4-5
OpenAI	gpt-4o, gpt-4-turbo, o1, o3
xAI	grok-3, grok-4
Google AI Studio	gemini-2.5-pro, gemini-2.5-flash
Vertex AI	gemini-2.5-pro, gemini-2.5-flash
Groq	llama-3.3-70b-versatile, mixtral-8x7b
Workers AI	@cf/meta/llama-3.2-3b-instruct

Default: Anthropic / claude-sonnet-4-5

Prompts

System Prompt (optional) Sets the AI’s role and behavior. Example:

You are a data analyst. Be concise and factual.

User Prompt (required) The task for the AI. Use template variables to reference data from upstream nodes:

Analyze these errors and prioritize by severity:
{{sentry.issues}}

Assistant Messages (optional) Add example responses to guide output format. Useful for few-shot prompting.

Output Format

Controls how the AI structures its response:

Format	Use Case	Configuration
Text	Summaries, explanations, chat	Default, no config needed
One result	Structured data extraction	Define JSON schema
Many results	Lists, batch processing	Define element schema
Classify	Categorization, routing	Define options list

Text (Default)

Free-form text response. Best for summaries, explanations, and conversational output.

One Result (Object)

Returns a single structured object. Define the schema:

{
  "type": "object",
  "properties": {
    "summary": { "type": "string" },
    "severity": { "type": "string", "enum": ["critical", "high", "medium", "low"] },
    "affectedUsers": { "type": "number" }
  },
  "required": ["summary", "severity"]
}

Many Results (Array)

Returns a list of structured objects. Define what each element looks like:

{
  "type": "object",
  "properties": {
    "title": { "type": "string" },
    "priority": { "type": "number" }
  }
}

Classify (Choice)

Model selects from predefined options. Good for routing and categorization:

critical
high
medium
low

Tools

Adding tools transforms a single LLM call into an autonomous agent that can reason and take actions.

System Tools

Tool	Description
Web Fetch	Fetch a URL and read its content as text
Web Search	Search the web and get result links
Sandbox	Execute bash commands, read/write files

MCP Servers

Connect Model Context Protocol servers to give the AI access to external tools:

Sentry (list issues, search events)
GitHub (create issues, read files)
Slack (send messages)
Custom MCP servers from your workspace

When tools are enabled, the agent will:

Analyze the task
Decide which tools to use
Call tools and observe results
Iterate until the task is complete

Advanced Settings

Temperature (0-2) Controls randomness. Lower = more deterministic, higher = more creative.

0: Deterministic (same input = same output)
1: Balanced (default)
2: Maximum creativity

Max Steps (1-50) Maximum iterations when tools are enabled. The agent stops when:

Task is complete
Max steps reached
Error occurs

Default: 20 steps

Behavior

Without Tools

Single LLM call:

Input → Prompt → Model → Output

With Tools

Autonomous agent loop:

Input → Prompt → Model → [Tool Call → Result]* → Output

The agent continues calling tools until it determines the task is complete or hits max steps.

Examples

Simple: Summarize Text

Model: claude-sonnet-4-5
Prompt: "Summarize this article in 3 bullet points: {{http.response}}"
Output: Text
Tools: None

Structured: Extract Data

Model: claude-sonnet-4-5
Prompt: "Extract contact info from: {{webhook.body}}"
Output: One result
Schema:
  name: string
  email: string
  phone: string (optional)
Tools: None

Agent: Research Task

Model: claude-opus-4-5
System: "You are a research assistant."
Prompt: "Find the latest pricing for {{input.competitor}} and summarize."
Output: One result
Tools: Web Search, Web Fetch
Max Steps: 10

Classification: Route Tickets

Model: claude-haiku-4-5
Prompt: "Classify this support ticket: {{zendesk.ticket}}"
Output: Classify
Options:
  - billing
  - technical
  - feature-request
  - other

Best Practices

Use text output for simple tasks - Don’t over-structure when plain text works
Use structured output for downstream processing - When other nodes need to parse the result
Enable sandbox only when needed - It grants filesystem access
Set reasonable max steps - 5-15 for most tasks, up to 50 for complex research
Use system prompts for consistent behavior - Define role and constraints upfront