The n8n AI Agent, explained: what the node does, when to use it, and 12 workflows to start from

What the AI Agent node actually is

n8n has two kinds of language-model nodes. Chains — Basic LLM Chain, Summarization Chain, Question and Answer Chain, Information Extractor, Text Classifier, Sentiment Analysis — make one model call and return. The AI Agent node makes a model call, lets the model choose a tool, runs the tool, feeds the result back, and loops until the model says it is done. The difference is the loop: a chain answers; an agent works.

The node is built from sub-nodes you attach to it: a chat model (Anthropic, OpenAI, Google, Mistral, Ollama and others), optional memory (Simple Memory for a session window, Postgres or Redis chat memory for persistence), any number of tools (HTTP Request Tool, Code Tool, Call n8n Workflow Tool, Vector Store Tool, Calculator, plus most n8n app nodes usable as tools), and an optional output parser that forces the final answer into a JSON schema so the next node gets typed fields instead of prose. The system message you write in the node’s options is where the agent’s job, rules and refusals live.

If you have read the agentic loops guide, this is that loop with n8n handling the observe and act sides. The model reasons; n8n observes (triggers, data pulls) and acts (tools, deliveries, approvals).

The one rule: deterministic first, agent second

Most of any business workflow is deterministic: pull the record, check the status, compute the number, send the message. Deterministic work belongs in plain nodes — Code, IF, Switch, Set, HTTP Request — because they are free, instant, testable and never hallucinate. The agent node earns its place only where the task requires reading language, making a judgement, or deciding which of several tools to use: classifying a ticket, extracting fields from a messy PDF, drafting a message from context, explaining an anomaly, deciding whether a request is complete.

Every blueprint on this site follows that rule. Look at the demand forecasting workflow: eight steps, one of which is an agent — and that agent explains the forecast, it does not produce it. Or payroll automation: the agent annotates anomalies and has no write access to amounts. Putting the model in the smallest box that still does the job is what makes these workflows reliable enough to sell.

A practical corollary: use the chain nodes (Text Classifier, Information Extractor) instead of the full agent for single-call jobs. They are cheaper, faster and produce structured output natively. Reserve the AI Agent node for steps that genuinely need tools or multi-step reasoning.

Building blocks that make agents safe in production

• Structured Output Parser. Attach it and set a JSON schema with a confidence and a needs_human field. Downstream IF nodes route on those, and the agent cannot return prose where a number is expected.
• Tools with narrow scope. An HTTP Request Tool that can only call one read endpoint is safer than a generic one. For write actions, do not give the agent the tool — route its decision to a Wait node with an approval button and let n8n perform the write after a human click.
• Call n8n Workflow Tool. Wrap complex or sensitive operations in a sub-workflow with its own validation, and expose that as the tool. The agent sees one clean function; you keep control of what it does.
• Error Trigger and an error workflow. Every production workflow should set an error workflow that logs the failure, alerts a channel, and — for idempotent steps — retries. Model APIs fail occasionally; the workflow should not.
• Idempotency keys. Store a key per processed item (ticket id, invoice number) and check it before acting, so a retried run cannot send the same email twice.
• Evaluations. Keep a sheet of 30–50 representative inputs with expected outputs and run the workflow against it whenever you change a prompt or a model. n8n’s evaluation features and a simple Compare Datasets node both work for this.

Where n8n agents stop and code begins

n8n is the right tool for the large majority of business automations that include a model. It is not the right tool for everything. Long-running agent loops with dozens of tool calls and complex state are hard to observe and debug in a visual canvas; multi-agent systems with supervisors and hand-offs get unwieldy; and rigorous evaluation and testing are easier in code. Teams building a product feature rather than an internal workflow usually reach for a framework — LangGraph, the Claude Agent SDK, the OpenAI Agents SDK — or plain code with the model’s API. The frameworks comparison covers those trade-offs; AI agent vs n8n covers the decision itself.

A useful test: if the workflow can be drawn on a whiteboard as a line with one or two decision points, it belongs in n8n. If the whiteboard needs a loop within a loop, or the agent has to plan a variable number of steps, prototype in n8n and expect to move to code if it sticks.

Cost control

Model usage is billed per token by the provider; n8n itself is free self-hosted (Sustainable Use License) or priced per execution on n8n Cloud. In practice the token bill for a business workflow is small and the mistakes that inflate it are predictable: sending the whole record to the model when it needs three fields; using the largest model for classification; letting an agent loop without a step limit; and re-processing items on retries. The fixes are workflow hygiene — Set nodes that shape the input, a smaller model for chain steps, a max-iterations option on the agent, and idempotency keys. The templates ship with those defaults.

Hosting and data

n8n Cloud is the fastest start and fine for most teams. Self-hosting (Docker on a small VM, or Kubernetes) is the choice when data must stay in your environment or when execution volume makes it cheaper; it also lets you run a local model through Ollama for sensitive steps. Credentials live in n8n’s encrypted store either way; the model provider sees only what the workflow sends it, which is one more reason to shape inputs with a Set node before the agent.