Agents

The Webhook node is your workflow's front door. Add it as the trigger, set HTTP Method to POST, and give Path a name you recognize. n8n then shows two URLs: a Test URL and a Production URL.

While you build, click "Listen for test event". n8n registers the Test URL and waits about 120 seconds for one incoming call, so you can watch the data land right in the editor. The Production URL only works after you save and activate the workflow, and those runs show up under the Executions tab, not on the canvas.

To answer the caller, set the Webhook's "Respond" parameter: "Immediately" returns a default 200 at once, "When Last Node Finishes" waits for the whole workflow, and "Using 'Respond to Webhook' Node" lets you shape the reply with a dedicated node at the end.

Your message arrives nested under body, so you read it as {{ $json.body.message }}. Paste the Test URL into the chat tool above, click "Listen for test event", send a message, and the execution should appear.

Add an AI Agent node and connect a Chat Model sub-node to it (OpenAI, Mistral, Ollama, whatever you hold a credential for). Since n8n 1.82 the Agent always runs as a Tools Agent, so the Chat Model connection is required: it is the brain the agent thinks with.

The whole game is the system prompt. Tell the model to reply with ONLY one category from a closed list: quote, support, application, spam. No sentences, no explanation, no punctuation. The tighter you constrain the output, the more stable it stays across messages.

Feed it the incoming text with an expression. Webhook data is nested under body, so reference the message as {{ $json.body.message }}. Run an easy one first (the quote), then throw the ambiguous and the spam messages at it and check it stays consistent.

Now ask the agent for structured output, not just a label. The clean way is to turn on "Require Specific Output Format" on the AI Agent and attach a Structured Output Parser sub-node.

In the parser, pick a Schema Type. "Generate From JSON Example" is the beginner-friendly one: paste a sample object like { "category": "", "name": "", "email": "", "subject": "" } and n8n infers the schema, treating every field as required. "Define using JSON Schema" gives you full control when you want it.

The golden rule: every field is always present, even when empty. Test with message 4 (the spam), where there is no name and no email. A well-built parser returns those fields as empty strings instead of dropping them, and that predictable shape is what lets the next nodes trust the data.

Add a Switch node after the agent and feed it the category field. In "Rules" mode you add one rule per output: category equals quote goes to the first output, support to the second, application to the third. Use "Rename Output" to label each branch, and turn on the "Fallback Output" to catch anything that matches nothing.

Think of each branch as a real action: a quote goes to sales, support opens a ticket, an application goes to people ops. You do not have to build those for real yet. One Edit Fields (Set) node per branch, writing a descriptive action field, is enough to prove the routing.

Send spam down its own branch that simply drops it. This is the moment your automation stops just reading and starts deciding.

There is no persistent memory in this workflow, so the simple trick is to let the model spot the signal in the text itself. In the system prompt add a rule: if the message mentions a previous contact, a reminder, or a chase-up, set a follow_up field to true.

Message 5 is a follow-up to message 1, same sender, same 200-unit quote. A good prompt flags it from cues like "I already wrote last week" and "following up".

This is a deliberate taste of a limitation. Real agents that link conversations over time need actual memory or a database to look the sender up. For now you are recognizing the pattern from language alone, which is enough to feel where the ceiling is.