You build an AI agent in n8n from four building blocks: trigger, chat model, memory and tools. The guide with practical tips and common mistakes.
You build an AI agent in n8n from four building blocks: a trigger that starts the agent, a chat model as the brain, a memory for conversation context, and tools with which the agent can actually do something. The AI Agent Node connects these parts, entirely without programming. This guide walks through building a first productive agent and shows the points where beginners typically get stuck. As of: July 2026.
Create a new workflow and choose the Chat Trigger as the starting point if you want to chat with the agent, or a webhook if other systems are meant to address it. Behind it, place the AI Agent Node. It is the control center: everything else is not chained in sequence but docked below the Agent Node, and this is what visually and logically distinguishes agents from normal n8n workflows.
Dock a model onto the agent's chat model connector, for example OpenAI or Anthropic, and store the credentials. Practical rule for choosing a model: a cheap model is enough for tool calls and routine answers, and the expensive class is only worthwhile for complex trade-offs. The model choice can be changed at any time later, so start cheap and measure.
Without memory, the agent forgets what was discussed after every message. Simple Memory is enough for initial tests, while Postgres Memory is recommended for production so the context survives a restart. The most common beginner mistake lies exactly here: the memory needs a session key that is unique per conversation partner. If the agent is used via a webhook, you must set this key yourself from the incoming data, otherwise conversations get mixed up or the history stays empty.
Only with tools does the chatbot become a digital employee. Almost anything n8n can do works as a tool: a database query, a search of your documents via a vector store, sending an email, a dedicated sub-workflow. Describe each tool in one sentence precisely enough that the model understands what it is for, because the agent decides whether to use it based on this description. For critical actions such as sending or deleting, build in human approval: n8n comes with human-in-the-loop patterns for this, and this is exactly what takes decision-makers' fear of autonomous AI away.
In the Agent Node, you store the system instruction: who the agent is, what it may do, what it should never do, and in which language and tone it responds. Write it like a work instruction for a new temporary staff member, concrete and with examples. Then test it in the chat window with real cases from your everyday work, not with made-up questions, and check the execution view: there you see every tool decision the agent makes and immediately recognize if a tool description was ambiguous.
Where AI agents really save time in operations and what a productive use case looks like is shown in our overview AI Agents. If you would rather have the agent built for you while keeping control: that is exactly what our Have an AI Agent Built at a fixed price is for.
Yes, in n8n an AI agent is created entirely visually: trigger, chat model, memory and tools are connected by drag-and-drop, and the logic sits in the system instruction in everyday language. Programming knowledge only helps with special cases, such as custom data formats or special interfaces.
The ongoing costs consist of the model calls and hosting. An internal agent with moderate use is often in the low double-digit euro range per month in model costs, while self-hosted n8n on a German server costs only a few euros in server costs. It only gets expensive with uncontrolled loops, which is why limits belong in from the start.
Yes, n8n connects local models via Ollama, so no text leaves your premises. Expect two limitations: local models need decent hardware, and they are noticeably more error-prone at tool calling than the large API models. For sensitive data, the compromise is often a European API model plus clear data rules in the system prompt.
A use case is suitable for an agent if the task requires real decisions between several tools, for example looking something up first and then answering or forwarding depending on the result. If the process always runs the same way, a classic automation with a single AI step is enough, which is faster, cheaper and easier to control.
Founder of NordFlux. Spent four years automating processes at enterprise scale at Dräger, and now brings that depth to the mid-market — pragmatic and with full data sovereignty.
Certifications
Overview page of the most common n8n AI Agent error patterns: prompt, rate limits, memory, and output parser issues at a glance, with links to the n8n docs.
Make or n8n for AI agents: comparing differences in setup, cost per run, and control over data.
Why an n8n agent forgets the chat history, how session keys work, and when Simple Memory or Postgres Memory is the right choice.
Connecting trigger, chat model, memory, and tools is step one; an agent that reliably decides and escalates in daily operation is step two. NordFlux builds AI agents on n8n, from the first version through to managed production operation, including system prompt tuning and error handling. In an initial conversation, we look at your use case and show what's still missing for a stable agent in your situation.
