n8n AI Agent

n8n's AI Agent node is one of the most powerful features in any automation platform. It implements a fully functional ReAct (Reason + Act) agent loop — Claude or GPT-4 decides what actions to take, executes them via connected tools, observes the results, and continues until the goal is achieved. Building AI agents in n8n requires no custom code for the agent loop itself; you configure the goal, tools, and model. This guide walks through building practical AI agents.

MJK Supplies · Jan 21, 2026 · 3 min read

ShareX in f ↗

How the n8n AI Agent Works

The AI Agent node in n8n operates a loop:

The agent receives a goal (from the system prompt) and an input (from the user message or trigger)
The model (Claude or GPT-4o) decides what action to take next
If a tool is needed, the agent calls the appropriate tool node
The tool executes and returns a result to the agent
The agent evaluates the result and decides the next action
This continues until the agent determines the goal is achieved or the maximum iterations are reached
The agent returns a final response

This loop enables genuinely autonomous multi-step tasks — researching a company (multiple search and web requests), qualifying a lead (CRM lookup + enrichment API + AI assessment), or processing a support ticket (lookup + knowledge base search + response generation) — all from a single agent configuration.

Configuring the AI Agent Node

Model selection: Choose your language model — Anthropic Claude (recommended for most tasks), OpenAI GPT-4o (strong for function calling), or other supported models.

System prompt: This is the agent's identity and goal definition. Write it carefully:

You are a lead qualification agent for [Company]. Your goal is to: 1. Look up the lead's company using the company lookup tool 2. Assess the company's fit against our ICP (50-500 employees, SaaS/tech industry, based in US/UK/AU) 3. Score the lead 0-100 and classify as high/medium/low priority 4. Return a structured qualification summary Only use tools when necessary. Stop when you have a complete qualification assessment.

Tools: Connect tool nodes that the agent can call. Common tools:

HTTP Request nodes (call any external API)
n8n's built-in CRM nodes (HubSpot, Salesforce)
Calendar nodes (check availability, create events)
Database nodes (query Postgres, Airtable)
Email nodes (send or read emails)

Memory: The agent can use memory nodes to maintain context across multiple messages (for conversational agents) or sessions (for agents that need to remember past actions).

Max iterations: Set a reasonable maximum (10-20 for most tasks). Without a limit, a confused agent can loop indefinitely.

Example: Lead Research Agent

This agent receives a company name and produces a full research brief.

System prompt:

You are a B2B sales research agent. Given a company name, research the company and produce a brief covering: 1. What the company does and their primary market 2. Company size and funding stage (if available) 3. Technology stack or key tools they use 4. Recent significant news or changes 5. Who the likely decision-makers are for [your product category] Use the web search and company lookup tools to gather this information. Stop when you have enough to write a complete brief.

Tools connected:

Web Search tool (Google or Serper API via HTTP node)
Apollo.io API (for company data via HTTP node)

Input: {{ $json.company_name }}

Output: The agent returns a complete research brief that the sales workflow routes to the appropriate CRM record.

Example: Customer Support Agent

System prompt:

You are a customer support agent for [Company]. Help the customer resolve their issue. You have access to: customer account lookup, order lookup, knowledge base search, and a refund tool (for amounts under $50). If you cannot resolve the issue with these tools, create an escalation ticket. Always be empathetic and solution-focused.

Tools connected:

Customer lookup (HTTP Request to internal API)
Order lookup (HTTP Request to order management system)
Knowledge base search (Vector store retrieval tool)
Issue refund (HTTP Request to payment system)
Create escalation (HTTP Request to ticketing system or HubSpot CRM)

Trigger: Webhook from your support email system

This agent handles 60-70% of support tickets completely autonomously, escalating only genuinely complex cases.

Debugging Agents

When an agent doesn't behave as expected, use n8n's execution view:

Click on a failed execution in the execution list
Click on the AI Agent node to see its full reasoning trace
Review each step: what did the agent decide to do? What did the tool return? Why did the agent make its next decision?

The full reasoning trace is visible in n8n — you can see every thought step and tool call, which makes debugging agent behaviour much more practical than debugging code.

Recommended Tools

n8n — Platform for AI agent workflows
Claude API — Recommended model for complex reasoning agents
OpenAI API — Strong alternative for function-calling agents
Apollo.io — Lead research data tool
HubSpot — CRM integration for sales and service agents
Pinecone — Vector store for knowledge base search

“The AI Agent node in n8n is the closest thing to autonomous AI in a visual automation platform. Invest time understanding how it works — it unlocks a category of use cases that aren't possible with standard workflow automation.”