Building AI Agents and LLM Workflows in n8n
n8n has become a practical orchestration layer for AI: connect any LLM, use the AI Agent nodes for tool calls, ground answers in your data with RAG and vector stores, and chain it all into real workflows. How to build it - self-hosted or on n8n Cloud - and the guardrails you actually need.
Table of contents
Chat interfaces are great for one-off questions. But the moment you want an LLM to do something repeatable — pull from your database, call an API, decide the next step, and run on a schedule — you need an orchestration layer. n8n has quietly become one of the most practical places to build that, because it combines a visual workflow canvas with dedicated AI Agent nodes and the ability to connect to essentially any LLM.
This is a tour of how you actually build AI agents and LLM workflows in n8n, and where the sharp edges are.
The LLM node and the AI Agent node
At the simplest level, n8n gives you an LLM node: feed it a prompt, get a completion back, pass that into the next node. You can wire in any model — you bring the credentials, n8n handles the plumbing. That alone covers a huge amount of ground: summarize incoming support tickets, classify leads, draft replies, extract structured data from messy text.
The more interesting piece is the AI Agent node. Instead of a single prompt-in, text-out call, an agent can use tools. You attach tools to the agent — an HTTP request, a database query, a search, another workflow — and the model decides which to call and when, looping until it has an answer. This is the difference between "the LLM writes text" and "the LLM takes actions," and it is the same shift from assistant to operator we described in AI Agents as Junior Developers: Where They Help and Where They're Dangerous.
Tool calls: giving the model hands
A tool in n8n is just a capability you expose to the agent. Want the agent to look up an order? Give it a tool backed by your database node. Want it to post to Slack, create a GitHub issue, or hit an internal API? Each becomes a tool the model can invoke, with n8n enforcing the actual execution.
This is conceptually close to the Model Context Protocol — a standard way to hand an AI assistant a controlled set of tools — which we broke down in MCP Explained Simply. n8n gives you a visual version of the same idea: the tools are nodes, the boundaries are explicit, and you can see exactly what the agent is allowed to touch.
RAG and vector stores
LLMs do not know your private documents. Retrieval-augmented generation (RAG) fixes that by fetching relevant chunks of your own data and feeding them to the model at query time. n8n supports this end to end: nodes to load and chunk documents, generate embeddings, store them in a vector store, and retrieve the closest matches to inject into the agent's context.
A typical RAG workflow looks like this:
- An ingestion workflow reads your docs, splits them into chunks, embeds them, and writes them to a vector store.
- A query workflow takes a user question, retrieves the most relevant chunks, and hands them to the AI Agent as context.
- The agent answers grounded in your data, optionally calling tools to take action on the result.
Because it is all visible on the canvas, you can inspect exactly what got retrieved and what the model saw — which makes RAG far less of a black box than it usually is.
Chaining steps into real workflows
Where n8n shines is combining these pieces with everything else it does. An agent step is just one node in a larger pipeline. You can trigger a workflow from a webhook, enrich the payload, run an AI Agent to decide what to do, branch on its output, call three different APIs, and write the result back — all in one inspectable flow. Splitting a big job across focused steps is the same discipline we covered in Subagents Everywhere: How to Split a Big Task Across Specialized Agents; n8n lets you express that structure visually instead of in orchestration code.
Own it or run it managed
You can build all of this on either a self-hosted instance or n8n Cloud. Self-hosting matters more than usual for AI workflows: your prompts, retrieved documents, and API keys stay on infrastructure you control, and there is no per-task metering on top of what you already pay your LLM provider. The tradeoff is the familiar one — self-hosting means you own updates, security, and uptime, the same ops reality covered in How to Host Your Own AI Agent (Without a DevOps Team). If you would rather not run a server, n8n Cloud is the managed path and the quickest way to get an agent running today.
The honest caveats
AI workflows fail in ways plain automations do not. Models hallucinate, tool calls misfire, and costs can creep if an agent loops more than you expected. Build in guardrails: cap the number of tool-call iterations, validate the agent's output before acting on it, log what the model did, and keep a human in the loop for anything destructive. The observability habits are not optional — they are what separates a reliable agent from a liability.
Do that, and n8n becomes a genuinely powerful orchestration layer: LLMs with tools, grounded in your data, wired into your real systems, running on your terms.
