The n8n AI Agent node with Claude: model + memory + tools in one drag-and-drop block

In n8n you can build a complete agent without writing code: drop in the AI Agent node, plug in a chat model (Claude), attach memory, add a few tools — and the agent decides which tool to call to finish the job. This article explains the structure, how to configure it with Claude, an order-lookup assistant example, and the common mistakes that make agents loop and burn tokens.

Structure: one root node, several sub-nodes

AI Agent is a “root node” with ports for sub-nodes:

Build steps

1. Add the AI Agent node and choose Tools Agent.
2. Connect an Anthropic Chat Model to the Model port; create a credential with a Claude Console API key and pick a model (Haiku for simple work, Sonnet/Opus when reasoning matters).
3. Write the System Message (standing rules) and the Prompt (usually an expression pulling data from the previous node, such as the chat message).
4. Attach tools to the Tools port.
5. Attach memory for multi-turn conversations.
6. Set Max Iterations and enable Return Intermediate Steps while debugging.

Example: an order-lookup assistant for customer support

Chat Trigger (support agent asks in an internal chat)
  → AI Agent (Tools Agent)
       ├─ Model:  Anthropic Chat Model (Sonnet)
       ├─ Memory: Postgres Chat Memory (per support-agent session)
       ├─ Tool:   Postgres "find_order" — look up by order ID or phone number
       └─ Tool:   HTTP Request "track_shipment" — call the carrier's API
  → reply in chat + log to Google Sheets

Sample System Message:

You are an order-lookup assistant for customer-support staff.
- Only answer questions about orders and shipping.
- Always look things up with tools; never guess a status.
- Never show a customer's full phone number or address; last 3 digits only.
- If unsure, say "not found" and suggest re-checking the order ID.
- Reply in at most 5 lines, with the order ID and latest status.

Write tool descriptions like docs for a colleague

The agent chooses tools from their descriptions. “Database tool” is close to useless. Compare:

Weak:   "Database tool"
Strong: "Find an order by order ID (format DH123456) or phone number.
         Returns: order ID, order date, status, tracking number.
         Use when staff ask about a specific order."

Rules of thumb: one job per tool; state inputs, outputs and when to use it. To reuse complex logic, package it as its own workflow and attach it with the Workflow Tool.

Common problems

Safety

• Write tools (deleting, refunding, emailing customers) belong behind a human-approval step — for example Slack “Send and Wait for Response”.
• Give tool credentials minimal rights: lookup tools use a read-only DB user.
• If the agent reads customer content (email, chat), treat it as untrusted input — it may contain prompt injection.
• Watch the Executions tab to spot loops and errors early.

How to apply it

• Developers: Start with 1–2 read-only tools, turn on intermediate steps to see what the agent “thinks”, and refine tool descriptions until tool choice is reliable.
• Tech Leads: Standardise a system-message template (role, limits, output format) and an approved tool list for internal agents.
• PMs: An n8n agent is the fastest way to test an internal-assistant idea in a few hours — prove the value before investing in code.