Phase 2: n8n for AI Automation · 55 min · n8n · OpenAI API · Anthropic API

AI Agent Nodes in n8n — LLMs, Tools, and Memory

Agents decide which tools to call; you define the tools.

Hiring signal: AI agent configuration is the skill that distinguishes automation engineers from automation users. Being able to configure an n8n AI Agent node with LLM, tools, memory, and iteration limits — and explain when to use agents vs fixed workflows — demonstrates the depth that production automation roles require.

What you will learn

Configure n8n's AI Agent node with LLM model, tools, memory, and system prompt
Define tools for agents: HTTP requests, database queries, sub-workflow calls
Configure conversation memory: window size, token buffer
Set maximum iterations to prevent infinite loops and control cost

The Problem

A customer support team builds an n8n workflow: receive ticket → LLM classifies category → IF node routes by category → separate workflow per category. It works for 80% of tickets. But the other 20% are ambiguous — a refund request that's also a complaint, a billing question that requires checking the order database, a technical issue that needs both knowledge base lookup and account history. The fixed IF/ELSE routing can't handle these. The team needs a system that can look at a ticket, decide what information it needs, call the right tools to get that information, and then respond.

This is what AI agents do. Instead of hardcoding the decision tree, you give the LLM a set of tools and let it decide which to call based on the input. n8n's AI Agent node makes this visual and production-ready.

An agent is not a smarter LLM call — it's a loop
The agent node implements a ReAct-style loop: the LLM receives the input, decides which tool to call, gets the tool's result, decides if it needs another tool, and repeats until it can answer. Your job is to define good tools and set boundaries (max iterations, memory, system prompt).

The Concept

AI Agent Node Anatomy

┌─────────────────────────────────────────────────────────────┐
│  AI Agent Node                                              │
│                                                             │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐         │
│  │  LLM Model  │  │   Tools     │  │   Memory    │         │
│  │             │  │             │  │             │         │
│  │ OpenAI      │  │ HTTP Req    │  │ Window      │         │
│  │ Anthropic   │  │ DB Query    │  │ Buffer      │         │
│  │ Ollama      │  │ Sub-workflow│  │ Token       │         │
│  │ Azure       │  │ Calculator  │  │ Buffer      │         │
│  └─────────────┘  └─────────────┘  └─────────────┘         │
│                                                             │
│  ┌─────────────────────────────────────────────────┐       │
│  │  System Prompt                                   │       │
│  │  "You are a support agent. Use tools to look up  │       │
│  │  order info before responding. Always cite the   │       │
│  │  order number in your response."                 │       │
│  └─────────────────────────────────────────────────┘       │
│                                                             │
│  Max Iterations: 5                                          │
└─────────────────────────────────────────────────────────────┘

The Agent Loop

LLM Providers Supported in n8n

Provider	n8n Node	Models	Best For
OpenAI	OpenAI Model	GPT-4o, GPT-4o-mini, o3-mini	Default choice, tool calling, fast
Anthropic	Anthropic Model	Claude 3.5 Sonnet, Haiku	Long context, careful reasoning
Ollama	Ollama Model	Llama 3, Mistral, Qwen	Local, no API cost, privacy
Azure OpenAI	Azure OpenAI Model	GPT-4o (Azure)	Enterprise, data residency
Google	Google Gemini Model	Gemini 2.0 Flash/Pro	Multimodal, long context

Defining Tools

Tools are the agent's capabilities. In n8n, you connect tool nodes to the AI Agent node:

Tool Type	n8n Node	Example Use
HTTP Request	HTTP Request Tool	Call external APIs, webhooks
Database Query	Postgres/MySQL Tool	Look up customer, order, inventory
Sub-workflow	Execute Workflow Tool	Run complex logic as a tool
Calculator	Calculator Tool	Math operations
Knowledge Search	Vector Store Tool	RAG retrieval
Custom	Code Tool	Any custom JavaScript logic

Each tool has:

Name: Short identifier (e.g., lookup_order)
Description: What the tool does (the LLM reads this to decide when to use it)
Input schema: What parameters the tool accepts (JSON Schema)

Tool descriptions are the most important prompt engineering in agent systems
The LLM decides which tool to call based on the description. If your description says "lookup" but the tool actually searches by email, the LLM will pass the wrong parameters. Write descriptions like documentation: what it does, what input it expects, what it returns.

Memory Configuration

Memory lets the agent remember previous turns in a conversation:

Memory Type	How It Works	Best For
Window Buffer	Keeps last N messages	Short conversations, simple
Token Buffer	Keeps messages until token limit	Cost control, long conversations
Postgres/Redis	Persists across sessions	Multi-turn chatbots, user history
No memory	Each call is independent	One-shot tasks, batch processing

You're building a support agent that handles multiple questions from the same user in a session. The agent keeps forgetting the user's order number from earlier in the conversation. What should you configure?

The agent isn't forgetting because it's not smart enough — it's forgetting because there's no memory configured, or the window is too small. Window Buffer memory keeps the last N messages in context. Increase the window size so the order number from earlier turns stays in the LLM's context window.

Maximum Iterations

The max iterations setting prevents infinite loops and controls cost:

Setting	Effect	Use Case
1–3	Agent can call 1–3 tools before stopping	Simple lookups, single-step
5 (default)	Agent can call up to 5 tools	Most support/automation tasks
10+	Agent can do multi-step research	Complex investigation, data gathering

Each iteration is an LLM call. At 5 iterations with GPT-4o-mini, a single agent run costs ~$0.003–$0.01. At 10 iterations with GPT-4o, it can cost $0.05–$0.15 per run.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Build It, Use It, Ship It, Exercises, Key Terms, Common Pitfalls — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.