n8n AI Agent nodeConfigure AI Agent in n8n.

This selector says where the agent picks up the user message. It is the first thing to check when the node runs on an empty prompt.

• Connected Chat Trigger Node (auto): reads the chatInput field from a Chat Trigger wired straight into the agent.
• Connected Guardrails Node (guardrails): reads guardrailsInput from a directly connected Guardrails node.
• Prompt (User Message): the required field opened by define (Define below), where you type text or an expression such as {{ $json.subject }}.

Switch this on when the answer has to be machine readable. The node then exposes the output parser port and stops returning a paragraph of prose.

• Require Specific Output Format: a boolean. On, the agent is expected to produce the shape described by the parser you connect; off, you get plain text.

A safety switch for the model side. Turned on, the agent is configured with a fallback model in addition to the main one, so a failing call has somewhere to go.

• Enable Fallback Model: a boolean, off by design on a simple agent, worth turning on when the workflow runs unattended.

Here you write the instruction sent to the agent before the conversation starts. Most of the quality of an agent lives in this field, not in the model choice.

• System Message: free text, sent ahead of the user message. Describe the role, the tone, what to do when a tool fails, and when to answer nothing rather than guess.

An agent works in loops: think, call a tool, read the result, think again. This number caps how many of those loops it gets before it stops.

• Max Iterations: the maximum number of iterations the agent runs before stopping. Raise it for a task with several lookups, lower it when you want a fast failure instead of a long spiral.

By default the node hands back only the final answer. Turn this on and the output also contains the steps the agent went through to get there.

• Return Intermediate Steps: a boolean. On, the output includes the tool calls and their results; off, the payload stays small.

When an item arrives with an image attached, this option hands that image to the agent as an image message rather than dropping it.

• Automatically Passthrough Binary Images: a boolean that decides whether binary images are passed through to the agent automatically.

This option attaches your own labels to the tracing events of the run, so an execution can be identified later by something other than its timestamp.

• Key: the name of the label, for instance a customer or a campaign.
• Type: the value type, among stringValue, numberValue, booleanValue, arrayValue and objectValue.
• Value: the content itself, often an expression like {{ $json.customerId }}.

Highlighted data is what lets you filter the Executions view on your own values instead of scrolling through a list of runs.

• Auto-save highlighted data: a boolean, on by default, available on the Pro and Enterprise plans of n8n Cloud, and on Enterprise or a registered Community Edition when you self-host.

Streaming sends the answer back as it is being written instead of waiting for the last word, which changes how a long generation feels to whoever is waiting.

• Enable Streaming: a boolean, enabled by default. It only works when the workflow starts from a trigger that supports streaming responses, a Chat Trigger or a Webhook node with Response Mode set to Streaming.

When a hundred items hit the agent at once, this is where you slow the node down on purpose to stay inside the rate limits of the model provider.

• Batch Size: how many items are processed in parallel. It helps against rate limits, and it can change the ordering of the log output.
• Delay Between Batches: a pause in milliseconds between two batches, for the same reason.