Complete Guide to Building AI Agents with n8n

An AI agent is a workflow that uses a language model to decide what to do next: answer, ask a follow-up question, or call a tool such as your CRM or knowledge base. n8n is a good place to build agents because the agent sits inside ordinary workflows, so you get triggers, error handling, logging and every integration n8n already has.

This guide goes from a simple chatbot to a multi-agent system, in the order we build them for clients.

How an n8n agent fits together

Every agent in n8n has the same parts:

• A trigger. The Chat Trigger node for a chat window, a Webhook for your own app, or an app trigger such as a new email or a new ticket.
• The AI Agent node. It holds the instructions (the system message) and runs the loop: read the input, decide, call tools, answer.
• A chat model. A sub-node for the model that does the reasoning, such as OpenAI, Anthropic or a self-hosted model.
• Memory. A sub-node that keeps the conversation, from simple memory for testing to Postgres or Redis for production.
• Tools. Sub-nodes the agent may call: other workflows, HTTP requests, vector store search or built-in app actions.

Before you add a single node, write the agent's job in one paragraph: who it serves, what it may decide on its own, which systems it may touch and when it must hand off to a person. That paragraph becomes the core of the system message and keeps the tool list short.

Step 1: a simple chatbot

The smallest useful agent is a chatbot that answers from a short brief.

1. Add a Chat Trigger. It gives you a hosted chat page for testing and a chat you can embed on your site.
2. Add the AI Agent node and connect a chat model.
3. Write the system message: role, tone, what to answer, what to decline and what to say when it does not know.
4. Add memory keyed to the chat session so follow-up questions make sense.

Test it with the questions real customers ask, including the awkward ones. If answers depend on facts that change, such as prices, policies or stock, do not paste them into the prompt. Give the agent a tool or a knowledge base instead.

Step 2: tool calling

Tools turn a chatbot into an agent. The most reliable pattern in n8n is one sub-workflow per tool, connected with the Call n8n Workflow Tool:

• Name and describe each tool clearly. The model picks tools by reading their descriptions. "lookup_order: find an order by order number and return its status and delivery date" works far better than "orders".
• Define the inputs the tool expects, and validate them in the sub-workflow before touching any system.
• Return small, clean results. The model needs the fields that answer the question, not the whole API response.
• Separate reads from writes, so you can put approvals in front of the risky actions.

For quick integrations, the HTTP Request Tool lets the agent call an API directly, and many app nodes can be attached as tools. Sub-workflows are still worth it for anything that writes data, because you control exactly what happens.

Step 3: RAG for answers grounded in your documents

Retrieval-augmented generation (RAG) lets the agent answer from your own documents instead of the model's general knowledge. It takes two workflows.

Ingestion loads your documents, splits them into chunks, turns each chunk into an embedding and stores it in a vector store. n8n has nodes for Pinecone, Qdrant, Supabase, Postgres with pgvector and others. Run ingestion on a schedule, or when a file changes in Google Drive, Notion or your help center, so answers stay current.

Retrieval happens at question time. Connect the vector store to the agent as a tool, so it searches when it needs facts and passes only the most relevant chunks to the model.

What makes RAG work in practice:

• Clean the sources first. Outdated or duplicate documents produce confident wrong answers.
• Chunk by meaning, such as one section or one FAQ per chunk, and keep titles and URLs as metadata so the agent can cite where an answer came from.
• Allow "I don't know". Tell the agent to say when the documents do not cover a question instead of guessing.

Step 4: multi-agent systems

One agent with twenty tools gets confused. When the job grows, split it:

• An orchestrator agent talks to the user and decides who should handle each request.
• Specialist agents each own one area, such as billing, scheduling or technical support, with their own instructions and a few tools.

In n8n, a specialist can be a sub-workflow with its own AI Agent node that the orchestrator calls as a tool, or an AI Agent Tool node attached directly to the orchestrator. Keep the handoff simple: the orchestrator passes a clear task with context, and the specialist returns a short, structured result.

Reach for multiple agents only when one agent with good tools is not enough. Each extra agent adds cost, latency and another place for things to go wrong.

Step 5: structured output and routing

Agents are most useful inside larger automations, where the next step needs predictable data. Add a Structured Output Parser, or ask for JSON that matches a schema, then route with Switch and IF nodes. A support agent can return a category, an urgency and a draft reply, and the workflow decides whether to send it, escalate it or open a ticket.

Guardrails before you launch

• Least privilege. Give each credential only the permissions its tool needs. A lookup tool does not need write access.
• Human approval. Pause before actions that spend money, message customers or change important records. n8n can send an approval request to Slack or email and wait for the answer.
• Untrusted input. Treat everything users type, and everything retrieved from documents, as untrusted. Validate tool inputs and never pass them straight into queries or commands.
• Limits. Cap message length, tool calls per conversation and daily model spend, and rate-limit public chat endpoints.
• Logging. Record every tool call with its inputs and result, so you can always explain what the agent did.

Testing and monitoring

Build a set of test questions with the answers you expect, including edge cases and requests the agent must refuse. Run it whenever you change the prompt, the model or a tool, and compare results before you publish. n8n's evaluation features can run these test sets inside the workflow.

After launch, review real conversations every week at first: where the agent failed, which tools it misused and what people asked that it could not answer. Most improvements come from clearer tool descriptions and cleaner documents, not from a bigger model.

Deploying and managing agents

• Keep a staging copy of each agent workflow and promote changes to production on purpose, not by editing live.
• Keep the system message with the workflow, and note what changed and why.
• Run busy agents in queue mode with workers, so long model calls do not hold up other workflows.
• Set an error workflow that alerts your team when an agent run fails.

Where to start

Pick one narrow job with a clear measure of success, such as answering order status questions or qualifying inbound leads, and ship it before adding more tools or agents.

If you want help, our AI agent development team designs, builds and runs agents like these on n8n, and chatbot development covers customer-facing chat. For a website chatbot as part of a wider marketing build, see AI chatbot development at Devma Agency.