The n8n AI Agent node: tools, memory and human review

#01How the AI Agent node works

The AI Agent node is a root node. You attach:

• A chat model (required).
• Tools (at least one is required, according to the AI Agent docs). A tool can be an app node such as Google Sheets or HubSpot, an HTTP Request tool, a call to another n8n workflow, or an MCP client.
• Memory (optional), so the agent remembers earlier messages in a conversation.
• An output parser (optional), for structured results.

The agent reads the prompt, decides whether to call a tool, reads the result, and repeats until it has an answer. The n8n docs note that since version 1.82.0 all AI Agent nodes work as a Tools Agent, and the older v1 agent is removed in n8n 3.0.

#02Agent or workflow?

Agents make more model calls than a fixed chain, because they reason about which tool to use. They're also harder to test. Use them where flexibility pays for that.

#03Designing tools for an agent

Each tool needs a clear name and description, because the model decides which to call from those. Good tool design:

• One job per tool. "Get campaign rows for a date range", not "Do anything with Sheets".
• Read-only first. Give the agent operations like Get and Search. Leave Create, Update and Delete out until you've added human review.
• Wrap complex logic in a sub-workflow and expose it with the Call n8n Workflow tool. You control exactly what happens; the agent just supplies inputs.
• Describe the inputs the tool expects, including formats such as ISO dates.

App nodes used as tools can have parameters filled by the model. The n8n docs call this AI tool parameters. Fix anything that shouldn't vary, such as which sheet or which HubSpot pipeline.

Our view: the best first agent for a marketing team is an internal Q&A assistant over your own campaign data, with no write access. It teaches the team how agents behave at low risk.

#04Where MCP fits

The MCP Client tool lets an agent use tools exposed by an MCP server, and the MCP Server Trigger lets you expose n8n workflows as tools to MCP clients such as Claude. Lesson 15 covers both.

#05Tools marketers attach to agents

#06Human review for tools

The n8n docs on human-in-the-loop for tools describe how to require approval before an agent runs a specific tool. When the agent decides to call that tool, the workflow pauses and sends a request through a channel such as Slack, Microsoft Teams, Gmail, Telegram or n8n Chat. The reviewer sees which tool and what parameters, then approves (the tool runs) or denies (it doesn't, and the agent is told).

The docs recommend it for irreversible actions, compliance needs, high-value decisions, and while you're building trust. You can apply it to all of an agent's tools or just selected ones. The review channel can differ from where the conversation happens.

#07Memory types

According to n8n's memory docs, Simple Memory stores a configurable window of chat history for a session and is the easiest start. For persistence across restarts or many users, there are nodes for Redis, Postgres, Zep and others. Agents can use memory; chains can't.

#08Agent failure modes

#09Testing an agent

Write ten questions with known answers, including two it should refuse or say it can't answer. Run them after every change to the system message, tools or model, and check the tool calls in the execution log, not just the final text. Lesson 17 shows how to turn this into an evaluation.

#10Worked example: Ledgerline's campaign assistant

Ledgerline is a fictional accounts-payable software company. Details are illustrative.

Ledgerline's marketing lead wanted answers like "what's running next week?" without opening four tools. The team built the read-only assistant above with a Sheets tool and two sub-workflow tools: one for GA4 top pages and one for HubSpot form submissions last week.

Early tests showed the agent answering a budget question from the calendar sheet's notes column, which wasn't reliable. They tightened the tool description to "campaign names, dates, owners and channels only" and added "If asked about budget, say you don't have budget data" to the system message.

A Slack posting tool came later, behind human review, so the lead could ask for "a summary for #marketing" and approve it before it posted.

Walkthrough

1. Add a Chat Trigger so you can talk to the agent in n8n.
2. Add an AI Agent node. Attach a chat model sub-node.
3. Attach Simple Memory with a short window.
4. Attach a Google Sheets tool configured to read your campaign calendar (read-only operation).
5. Attach a Call n8n Workflow tool that runs a sub-workflow returning last week's top pages from GA4 (lesson 14 builds it).
6. System message: "You are Ledgerline's marketing assistant. Answer questions about campaigns and recent traffic using the tools. If the tools don't contain the answer, say so. Never guess numbers."
7. Ask: "What's scheduled next week, and which page had the most sessions last week?" Check the agent's tool calls in the output.
8. Add a Slack tool for posting summaries, behind human review.

#11Try it yourself

Build the campaign assistant above with read-only tools. Ask it five questions, including one it can't answer from its tools. Check that it says so instead of inventing an answer. Then add one write tool with human review and test both approve and deny.

#12Quick quiz

1. What must every AI Agent have, besides a model?

Answer: At least one tool sub-node.

2. What happens when a reviewer denies a tool call?

Answer: The action is cancelled and the agent is informed of the rejection.

3. Can a Basic LLM Chain use memory?

Answer: No. In n8n, agents can use memory but chains can't.

#13What should you check before the next lesson?

• Your first agent has read-only tools.
• Write tools are behind human review.
• The system message says what to do when unsure.

Next: An AI content pipeline with human review. Or return to the n8n for Marketers track.

#14Common questions

Is an agent better than a workflow? Not usually. Agents suit open questions; fixed processes are more reliable as workflows.

Can the agent post to Slack on its own? It can if given the tool. Put human review on any tool that sends or changes things.

Does the agent remember past chats? Only with a memory sub-node, and only within that memory's scope.