Question 1

How do you choose between an AI agent and a chain in n8n?

Accepted Answer

Explain the difference, then frame the choice around cost and predictability. "A chain runs a predetermined sequence: input arrives, the model is called, a result comes out. One call, one result. Its important limitation is that chains in n8n do not support memory — they cannot remember previous queries. An agent is the decision-maker: it decides which tool to call, evaluates the result and calls again if needed. So a single execution involves several model calls. My criterion is simple: if I know what should happen, use a chain. Classification, summarisation, field extraction, translation — none of these need an agent, and it only adds cost. If the model must decide, use an agent. Answering "what is this customer's situation?" might require the CRM, the payment system or the documents — and I do not know which in advance. The mistake I see most is putting an agent on everything: the cost multiplies and the results get less consistent."

Question 2

Explain the cluster node structure in n8n.

Accepted Answer

Describe the visual structure — the fastest way to show you understand it. "A cluster node is a root node with sub-nodes attached beneath it. The root node sits in the main data flow — input in, output out. The sub-nodes are not in the flow; they are resources the root node reaches for while executing. In the AI Agent example the root node is the agent itself, and the sub-nodes are: Chat Model — which language model it uses Memory — the conversation history Tool — what it can call; at least one must be connected Building RAG adds Embeddings, Vector Store and Document Loader to that list. The practical meaning: reading a sub-node as 'the next step' is wrong. They are not parallel paths but capabilities the root node calls on when it needs them."

Question 3

How would you control the cost of an AI workflow?

Accepted Answer

This is a senior question. List concrete levers. "1. A chain instead of an agent. The biggest saving is here. An agent calls the model several times per execution; for a fixed task a chain calls it once. Checking whether the task genuinely needs decisions is step one. 2. Reducing what reaches the model. Instead of passing the whole item, I keep only the needed fields with Edit Fields. Billing is per token, so every unnecessary field costs money. 3. Model tiering. A smaller, cheaper model for simple tasks and a stronger one for hard decisions. In n8n you branch with a Switch into different Chat Model sub-nodes. 4. Filtering early. I drop unwanted items with a Filter before they reach the model — spam and auto-reply emails never go to the AI at all. 5. Reusing results. When the same input yields the same result, I cache it rather than calling the model a second time. n8n also has evaluation features for measuring output quality. That is not a direct cost lever, but it answers "would quality drop if we moved to a cheaper model?" Changing models without measuring is a guess."

Aspect	Chain	Agent
How it works	Calls components in a predetermined sequence	The language model decides which action to take
Tools	None	Yes — at least one tool must be connected
Memory	Not supported — it cannot remember previous queries	Supported
Number of runs	Once	Several times — it calls a tool, evaluates the response and continues
Predictability	High — you know what will happen	Lower — the model decides
Cost	One model call	Several calls — more expensive

AI components in n8n: chains and agents

📚 Sources and documentation