Agent Node

Run a multi-step AI agent with access to tools and MCP connections

Last updated 8 Oct 2026

The Agent node runs an AI agent that can use built-in tools and connected MCP servers to complete multi-step tasks. Unlike the LLM node, the Agent node iterates in a reasoning loop — calling tools, observing results, and continuing until it produces a final answer (or hits a guardrail).

When to Use

Use Agent when a task requires multiple steps, tool calls, or decision-making that can't be expressed as a single prompt — for example: researching a company, drafting a personalized email, and then revising it based on a compliance check.

Configuration

The agent runs once per input record, resolving {{field}} placeholders against that record.

Field	Type	Description
`provider`	`openai` \| `anthropic` \| `google` \| `groq`	LLM provider (default: `anthropic`)
`model`	string	Model ID
`credentialId`	string	LLM credential to use. Leave blank to use the platform-managed key (subject to your plan)
`systemPrompt`	string	Agent instructions. Supports `{{variable}}` syntax
`userPromptTemplate`	string	Per-record user message. Supports `{{variable}}` syntax
`outputColumn`	string	Field to write the agent's final response to (default: `agent_output`)
`outputMode`	`text` \| `json_schema`	Whether to return free text or schema-validated JSON
`outputSchema`	object	JSON Schema for structured output (used when `outputMode: json_schema`)
`reasoningType`	string	Reasoning strategy that wraps the system prompt with chain-of-thought / structured-process guidance (`direct` is a no-op)
`builtInScopes`	string[]	Built-in MCP tool scopes to enable (e.g., `["census:read", "http:use"]`)
`builtInToolNames`	string[]	Restrict to specific built-in tool names (a filter within `builtInScopes`, not an additive source)
`externalConnectionIds`	string[]	IDs of external MCP connections available to the agent
`workflowToolIds`	string[]	Automation IDs to expose to the agent as callable tools
`maxIterations`	number	Maximum reasoning iterations per record
`maxToolCalls`	number	Maximum total tool calls per record
`costCapPerRow`	number	Maximum estimated LLM cost in USD per record before aborting
`timeoutPerRow`	number	Per-record timeout
`continueOnError`	boolean	If true, failed records are emitted with an error field rather than failing the run (default: `false`)
`trustLevel`	string	Governance trust level applied to the agent's tool calls

Example

Research and summarize each company:

provider: anthropic
model: claude-sonnet-4-6
systemPrompt: "You are a B2B research analyst. Research companies thoroughly and provide structured summaries."
userPromptTemplate: "Research {{company_name}} ({{domain}}). Summarize: industry, key products, recent news, and competitive position."
maxIterations: 15
costCapPerRow: 0.25
builtInScopes: [census:read, http:use]
outputColumn: company_research
continueOnError: true

Notes

Agent tool calls run on the governed MCP rail — guardrails are evaluated and the run is logged to the audit trail.
costCapPerRow is an approximate guard based on token estimation — actual costs may vary slightly.
Once maxIterations (or maxToolCalls) is reached, the agent stops and returns its best answer so far.
External MCP connections must be configured under Settings → Integrations before they can be referenced in externalConnectionIds.
Leaving credentialId blank uses the platform-managed provider key; frontier models may require a paid plan.

Node Reference Complete reference for all automation node types, organized by category LLM Node Generate text or structured data using a language model MCP Node Call a tool on a connected MCP server and map the response back to records Template Syntax Variable interpolation syntax for LLM, Agent, HTTP, and MCP node configurations

Still stuck? Browse the FAQ or contact us.

When to Use

Configuration

Example

Notes

Related articles