How to Build an AI Agent with n8n (2026 Guide)

How AI Agents Differ from Standard LLM Nodes in n8n

n8n has two separate nodes for AI work: the Basic LLM Chain node and the AI Agent node. Understanding the difference determines which one to use.

Basic LLM Chain node:

• Takes input, sends it to an LLM, returns the response
• Single round-trip: one prompt in, one response out
• No tool use, no iteration, no decision-making
• Use for: text transformation, summarisation, classification, content generation

AI Agent node:

• Gives the LLM a list of tools and a task description
• The LLM decides which tools to call and in what sequence
• Loops internally until the task is complete or the step limit is reached
• Use for: research tasks, multi-step data processing, customer support bots, autonomous workflows

The AI Agent node uses the ReAct (Reasoning and Acting) framework internally. The model produces a Thought, selects an Action (a tool), observes the Action Result, and repeats until it has enough information to produce a Final Answer.

[Thought] I need to search for current pricing on this product.
[Action] web_search("product name pricing 2026")
[Observation] Search returned 3 results with pricing data...
[Thought] I have enough data. I can now summarise.
[Final Answer] Based on the search results, the current price is...

This loop runs inside the AI Agent node — you do not need to wire up the loop manually in n8n.

When to Use Each Node

Set Up LLM Credentials

Option A: OpenAI (GPT-4o)

1. In n8n, go to Settings > Credentials > New Credential 2. Search for "OpenAI" 3. Paste your OpenAI API key 4. Name it "OpenAI account" and save

GPT-4o is the recommended model for complex agents — it has strong tool-calling performance and handles multi-step reasoning well.

Option B: Ollama (Local Models)

If you prefer local inference with no API costs, n8n supports Ollama via the "Ollama" credential type.

1. In n8n Settings > Credentials > New Credential, search for "Ollama" 2. Set the Base URL to your Ollama endpoint: - Local n8n instance: `http://localhost:11434` - n8n in Docker on the same machine as Ollama: `http://host.docker.internal:11434` - n8n on a different server: `http://[your-ollama-server-ip]:11434` 3. Save the credential

Verify the Connection

Create a temporary workflow with a single "OpenAI" or "Ollama" node, send a simple prompt ("Say hello"), and confirm you get a response. Delete the test workflow once confirmed.

Agent 1: Web Research Agent

This agent accepts a topic, searches the web, and returns a structured summary with sources. It uses two tools: a web search tool and a webpage reader.

Step 1: Create a New Workflow

Open n8n, click "New Workflow", and name it "Web Research Agent".

Step 2: Add a Manual Trigger (for testing)

Add a "Manual Trigger" node. This lets you test the workflow by clicking "Test Workflow" without needing an external trigger.

Step 3: Add the AI Agent Node

1. Click the "+" button to add a node after the trigger 2. Search for "AI Agent" and select it 3. In the node settings: - Chat Model: Select your OpenAI or Ollama credential - Model: `gpt-4o` or `llama3.3:8b` - System Message: Add the agent's role and instructions

Use this system message:

You are a research assistant. When given a topic, search the web for recent and relevant information, read the most relevant pages, and provide a structured summary with:
1. A 2-3 sentence overview of the topic
2. 3-5 key findings or data points
3. The sources you used

Always cite your sources with URLs. Be specific and factual. Do not include information you did not find in your search results.

Step 4: Add the SerpAPI Tool (Web Search)

1. Inside the AI Agent node, click "Add Tool" 2. Select "SerpAPI" (requires a free SerpAPI account — 100 searches/month free) 3. Add your SerpAPI credential 4. Set the description: "Use this tool to search the web for current information on any topic"

Alternative: Use the "Brave Search" tool if you have a Brave Search API key, or the "Tavily" tool for AI-optimised search results.

Step 5: Add the HTTP Request Tool (Webpage Reader)

1. Add another tool inside the AI Agent node 2. Select "HTTP Request Tool" 3. Configure it to fetch webpage content: - Method: GET - Description: "Use this tool to read the full content of a webpage when you have its URL" - The agent passes the URL parameter automatically

Step 6: Add an Output Node

Add a "Set" node after the AI Agent node to extract the agent's final output and format it for downstream use.

Step 7: Test the Workflow

Click "Test Workflow". In the Manual Trigger, add test data:

{
  "input": "What are the most popular AI coding tools in 2026 and how much do they cost?"
}

The AI Agent node shows each step in its output panel: the tool calls it made, the intermediate results, and the final answer.

Agent 2: Customer Support Agent with Memory

This agent handles multi-turn conversations, remembering what was said in previous messages. It uses n8n's Window Buffer Memory node to maintain context across turns.

Memory Node Types in n8n

For a customer support bot, Postgres Chat Memory is best — it persists across server restarts and keeps each user's conversation separate via a session ID.

Step 1: Set Up the Chat Trigger

1. Add a "Chat Trigger" node (instead of Manual Trigger) 2. Enable "Public" so the chat interface is accessible via a URL 3. The Chat Trigger automatically provides a `sessionId` for each user

Step 2: Add the AI Agent Node

1. Add an AI Agent node 2. Set the system message for a support agent persona:

You are a helpful customer support agent for AI Tool Discovery (aitooldiscovery.com). You help users find the right AI tools for their needs.

When a user describes what they want to do, ask clarifying questions to understand their specific use case, budget, and technical skill level. Then recommend 2-3 specific tools with brief explanations of why each fits their needs.

You remember previous messages in this conversation and refer back to them when relevant.

Step 3: Add Postgres Chat Memory

1. Click "Add Memory" inside the AI Agent node 2. Select "Postgres Chat Memory" 3. Connect your Postgres credential (the same database n8n uses works fine) 4. Set "Session ID" to: `{{ $('Chat Trigger').item.json.sessionId }}` 5. Set "Context Window Length" to 10 (remembers last 10 message pairs)

Step 4: Wire Up and Test

Connect Chat Trigger > AI Agent. Open the chat URL from the Chat Trigger node. Send several messages in sequence:

User: I want to automate my social media posting
Agent: [asks about platforms, frequency, content type]
User: Instagram and LinkedIn, once a day, mostly marketing content
Agent: [recommends specific tools, remembers Instagram + LinkedIn preference]
User: What was the first thing I told you?
Agent: [correctly recalls "automate my social media posting"]

The memory node stores and retrieves context automatically. The agent does not lose the conversation thread between turns.

Agent 3: Spreadsheet Data Processing Agent

This agent reads rows from a Google Sheet and decides what to do with each one based on its content. It demonstrates agents handling structured data with conditional decision-making.

Use case: A Google Sheet contains customer inquiry emails (copied in as text). The agent reads each row, classifies the inquiry type (refund, technical support, sales, general), and writes the classification back to the sheet.

Step 1: Schedule Trigger

Add a "Schedule Trigger" node set to run every hour (or on-demand for testing).

Step 2: Google Sheets Read

Add a "Google Sheets" node:

• Operation: Get Many Rows
• Spreadsheet: your target sheet
• Sheet: "Inquiries"
• Filters: rows where "Status" column is empty (unprocessed rows only)

Step 3: Loop Over Items

Add an "Item Lists" > "Split In Batches" node to process one row at a time. Set batch size to 1.

Step 4: AI Agent Node

Add the AI Agent node with this system message:

You are a customer inquiry classifier. You will receive the text of a customer inquiry.

Classify it into exactly one of these categories:
- REFUND: Customer wants a refund or money back
- TECHNICAL: Customer has a technical problem or bug report
- SALES: Customer is asking about pricing, upgrades, or new features
- GENERAL: Everything else

Respond with only the category name in capitals. No explanation, no punctuation, just the category.

Set "Input" to: `{{ $json.inquiry_text }}`

For this classification task, a smaller, faster model is appropriate. Use `gpt-4o-mini` (OpenAI) or `qwen2.5:7b` (Ollama) — classification does not need the full reasoning capability of the largest models.

Step 5: Google Sheets Write Back

Add a second "Google Sheets" node:

• Operation: Update Row
• Update the "Category" column with the AI Agent's output
• Update the "Status" column to "Processed"
• Match row by ID from the trigger data

Step 6: Error Handling

Add an "Error Trigger" node to catch failures. If the AI returns an unexpected value (anything other than the four categories), route it to a Slack or email notification for manual review.

AI Agent Tools Available in n8n

n8n's AI Agent node supports these built-in tool types as of n8n 1.25.x:

Custom Tools via HTTP Request

The HTTP Request tool is the most flexible. Configure it to call any external API — your CRM, your database API, your company's internal endpoints. The agent passes parameters you define, and you control the response format.

Example: a tool that checks order status by order ID:

Tool Name: check_order_status
Description: Use this tool to look up the current status of a customer order. Input: order_id (string)
URL: https://your-api.com/orders/{{ $fromAI('order_id') }}
Method: GET
Authentication: Header Auth (your API key)

The `$fromAI('parameter_name')` expression tells n8n to let the AI agent fill in that value based on what it knows from the conversation.

Limiting Token Usage

Agent loops consume more tokens than single LLM calls. To control costs:

- Set Max Iterations to the minimum needed (5-8 for simple tasks, 10-15 for research)
- Use smaller models for classification tasks (gpt-4o-mini instead of gpt-4o)
- Set a token budget in the system message: "Use a maximum of 3 tool calls total"
- Monitor usage in OpenAI's usage dashboard or your Ollama logs

Using Ollama Local Models as the Agent Backend

Swapping OpenAI for Ollama in an n8n AI Agent requires one credential change — everything else stays the same.

Recommended Models for n8n Agents

Changing the Backend in an Existing Agent

1. Open the AI Agent node 2. Click the Chat Model field 3. Switch the credential from "OpenAI account" to your "Ollama" credential 4. Change the model name to the Ollama model (e.g. `qwen2.5:14b`) 5. Save and re-test

No other changes are needed. The agent logic, tools, and memory configuration are model-agnostic.

Known Limitations with Ollama

Models smaller than 7B parameters frequently fail to output valid tool-call JSON, causing the agent loop to stall. If the agent stops responding mid-task or produces raw JSON in its output instead of calling a tool, switch to a larger model.

Ollama does not support streaming in n8n's Ollama Chat Model node as of n8n 1.25.x — responses appear all at once after the agent finishes processing. This is normal behaviour, not an error.

Troubleshooting

Frequently Asked Questions