From rigid automation to an agent chain

So far you have built automations that always run the same path. Trigger fires, step 1, step 2, output. Every branch you have to click in beforehand. That works as long as the world sticks to your plan. It rarely does.

An incoming email can be an invoice, a job application, a complaint or spam. A rigid chain treats all four the same, because it can only do what you clicked in for it. An agent reads the email, decides what it is and picks the next step itself. That is the jump we look at here, and it is smaller than you think.

The difference in one sentence

A classic automation knows the path. An agent knows the goal and finds the path.

With Zapier or a normal n8n workflow you set the logic. "If the subject contains X, then to Notion." You are the thinker, the tool is the hand. With an agent you give a goal and a few tools, and the model decides afresh on every run which tool it needs. You are the client, the agent is thinker and hand at once.

When the rigid chain is enough

The honest answer: more often than the AI world wants you to believe. If your process is unambiguous, stay with the fixed chain. A new Stripe payment always triggers the same confirmation email. There is nothing to decide there. An agent would be more expensive, slower and less reliable here, without a single advantage.

A rule of thumb I keep telling myself: if you can write the decision tree down in three sentences, you do not need an agent. Write it down and click it into n8n. Done.

When you really do need an agent

As soon as the right reaction depends on the content and the content is unpredictable. Three typical cases.

First, classification with fuzzy boundaries. "Is this support request urgent?" cannot be solved with a keyword filter, because "my shop is at a standstill" contains not a single urgency word and still has top priority. A model understands that, an if node does not.

Second, several steps whose order depends on the intermediate result. The agent researches, notices a piece of information is missing, fetches it, and then carries on. You cannot pre-build that loop in a fixed chain, because you do not know how often it runs.

Third, free language in, structured data out. Pulling the delivery address, the desired date and the article number out of an informal email. That is where an agent shines and a regex weeps.

What that looks like concretely in n8n

n8n has its own building block for this, the AI Agent node. You find it among the so-called cluster nodes. The principle: one root node, the AI Agent, and you hang smaller sub-nodes off it.

You connect three things to the agent. A language model, so OpenAI, Anthropic or a local model via Ollama. Optionally a memory node, so the agent remembers earlier parts of the same conversation. And at least one tool. That is mandatory, no tool no agent. The n8n docs say it literally, you have to attach at least one tool sub-node.

A tool is nothing magical here. It is a node the agent may call when needed. A Google search, an entry in a sheet, an HTTP call to some API. So you do not give the agent the sequence, you give it the toolbox. Which tool when is its decision.

Since version 1.82.0 n8n has simplified this. You used to be able to choose different agent types, that is gone. Now every AI Agent works as a so-called Tools Agent, which is exactly the toolbox principle. One less worry during setup.

The step beginners skip, and regret

An agent that decides freely can also be freely wrong. With a rigid chain you always know what happens. With an agent you have to deliberately take the control back.

n8n lets you require human approval before certain tools. The Tools Agent can be configured to stop before executing a delicate step and wait for your OK. Exactly what you built in the capstone as human in the loop, only one level up. My urgent advice: build that stop in before the agent is allowed to do anything irreversible. Sending an email, triggering a payment, deleting an entry. Better to ask one time too many than to send one time too many.

What that costs you, honestly

Every agent run calls the language model, often several times per task because it thinks in loops. That costs. A rigid chain with a single LLM call is considerably cheaper than an agent that thinks five times. Keep that in mind when you consider whether you really need an agent. Often a single well-built prompt in a fixed chain does the job, and you save money and nerves.

What you can do now

Take your capstone automation from the last lesson and ask yourself at exactly one place: am I making a decision here that depends on the content? If yes, replace that one if node experimentally with an AI Agent node with a single tool. Do not rebuild the whole workflow, just that one place. That way you feel the difference without getting lost.

The next level is about exactly this. From level 4 onwards it is about memory and MCP, so how an agent knows beyond a single run what it has already done and which tools it can reach. The AI Agent node here is the first small taste of that.

Source

• n8n AI Agent node, official docs: https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent (mandatory tool, Tools Agent from v1.82.0)
• n8n Tools Agent, detail including human approval: https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/n8n-nodes-langchain.agent/tools-agent
• LangChain in n8n, cluster node principle: https://docs.n8n.io/advanced-ai/langchain/overview