An n8n AI agent is what you get when a workflow stops following a fixed script and starts making its own decisions mid-run. Instead of “do step 1, then step 2, then step 3,” you give the agent a goal and a set of tools, and it figures out which tool to call and in what order.
This guide walks through what actually makes an n8n AI agent different from a normal workflow, the pieces every working agent needs, and how to build your first one — based directly on n8n’s own documentation, not just general AI-agent theory.
What Makes an n8n AI Agent Different?
According to n8n’s own node documentation, the AI Agent node lets you connect a chat model and one or more tools, and the agent decides which tools to call to complete a task. n8n defines an AI agent as an autonomous system that receives data, makes rational decisions, and acts within its environment to achieve specific goals — where “environment” means everything the agent can reach that isn’t the agent itself.
One specific requirement worth knowing upfront: you must connect at least one tool sub-node to an n8n AI agent for it to function as an agent at all. Without a tool attached, it behaves like a plain chatbot node rather than something that can actually act.
The Building Blocks: Chat Model, Tools, and Optional Memory
An n8n AI agent needs a connected chat model and at least one tool. Memory is optional, but it becomes important the moment the agent needs to remember context across a multi-message conversation rather than treating every message as a fresh, isolated request.
Chat Model
This is the LLM doing the actual reasoning — n8n supports connecting models like OpenAI’s GPT models, Anthropic’s Claude, and others through dedicated Chat Model sub-nodes. The model reads the task and decides what to do next.
Tools
Tools are what let the agent act instead of just talk. n8n’s official node library includes ready-made tool sub-nodes such as the Calculator, Custom Code Tool, MCP Client Tool (for calling any external MCP server’s tools), Wikipedia, Wolfram|Alpha, and the Call n8n Workflow Tool for using another workflow as a sub-agent — alongside any standard app node or HTTP request. The agent decides on its own which tool fits the current step; you don’t hard-code that logic.
Memory
Without memory, an n8n AI agent forgets everything between messages in the same conversation. n8n’s official tutorial demonstrates this directly: ask the agent your name, then ask “what’s my name?” — without memory attached, it can’t answer. Adding a Simple Memory sub-node fixes this by using the memory of the n8n instance itself, which n8n’s docs describe as usually sufficient for straightforward use cases.
Simple Memory does not persist between separate sessions. For memory that survives a restart or needs to scale across many concurrent conversations, n8n’s official node library also includes Postgres Chat Memory, Redis Chat Memory, and MongoDB Chat Memory nodes, each using that database as a dedicated memory server. A detail confirmed directly in both the Postgres and Redis Chat Memory docs: if you add more than one of either node to a workflow, they all access the same memory instance by default — worth knowing before running a destructive “override all messages” operation.
Step-by-Step: Building Your First n8n AI Agent
Step 1: Add a Trigger
Start with a Chat Trigger node if you want to interact with the agent conversationally, or a Webhook node if another system will call it programmatically.
Step 2: Add the AI Agent Node
Search for “AI Agent” in the node panel and add it to your canvas, connected to your trigger.
Step 3: Connect a Chat Model
Click the Chat Model connector at the bottom of the AI Agent node and choose your model provider. Without this connection, the agent has no way to reason about input.
Step 4: Connect at Least One Tool
Click the Tools connector and attach a first, low-risk tool — the Calculator tool is a safe starting point. This step isn’t optional: n8n’s own documentation requires at least one tool sub-node connected for the node to function as an agent at all, so it belongs before testing, not after.
Step 5: Test the Basic Response
Send a test message through the chat window to confirm the model reasons correctly and can call your connected tool.
Step 6: Add Memory Only If the Use Case Needs Multi-Turn Context
If your agent only needs to handle single, independent requests, you can stop here. If it needs to hold a conversation across multiple messages, click the Memory connector and select Simple Memory, then test a two-message exchange to confirm the agent now remembers what was said earlier.
Requiring Human Approval Before Risky Actions
For tools that perform sensitive actions — sending messages, modifying records, or deleting data — n8n’s Tools Agent documentation confirms you can require human approval before the agent executes that specific tool. This is configured directly on the tool connector in the Tools Panel, and it’s worth using deliberately: an n8n AI agent that can reason about which tool to call doesn’t automatically mean every tool it can reach should run unsupervised.
A Note on the Old Conversational Agent
If you find older n8n tutorials or templates referencing a “Conversational Agent,” here’s the current, precise status: it’s deprecated as of n8n 1.82.0, and all AI Agent nodes now work as a Tools Agent instead. Only workflows still using version 1 of the AI Agent node can still select the Conversational Agent — and that node version is scheduled for removal entirely in n8n 3.0, at which point it stops working for every workflow. If you’re building new, there’s no reason to reach for it; if you inherited an older workflow using it, plan to migrate before upgrading to n8n 3.0.
Common Pitfalls When Building an n8n AI Agent
A few mistakes show up repeatedly in early builds. The first is skipping memory entirely and being confused when the agent “forgets” context mid-conversation — this is expected behavior without a memory sub-node connected, not a bug.
The second is connecting too many tools at once. An n8n AI agent with a dozen loosely related tools attached tends to pick the wrong one more often than an agent with three or four well-scoped tools for a specific job. Start narrow, confirm reliable behavior, then expand.
The third is skipping the human-approval step for tools that can take irreversible action. It’s tempting to let an agent run fully autonomously once it’s working in testing, but testing rarely covers every edge case a real user will hit — a tool that deletes records or sends external messages is exactly where the approval gate earns its keep.
A Realistic Use Case
A common first project: an internal support-ticket triage agent. The trigger is a webhook from your helpdesk. The AI Agent reads the ticket text, uses a database tool to check if the customer has an active subscription, and uses an HTTP tool to post the categorized ticket into the right Slack channel. Memory isn’t critical here since each ticket is independent — but if you extended this into a multi-turn support chatbot instead, Simple Memory (or a persistent option for longer-term recall) becomes essential.
Notice what didn’t need to be hard-coded: the decision about which tool to check first, or how to phrase the Slack notification based on urgency, is left to the agent’s own reasoning rather than a chain of if/else conditions. That’s the actual value an n8n AI agent adds over a standard workflow for this kind of task — not speed, but the ability to handle tickets that don’t fit a predictable template.
n8n AI Agent vs a Standard n8n Workflow
A standard workflow executes a fixed sequence: node A always leads to node B. An n8n AI agent replaces that fixed sequence with a reasoning loop — the same LLM-driven decision-making covered in our n8n overview. Use a standard workflow when the logic is predictable and repeatable. Reach for an AI agent when the right next step genuinely depends on judgment a rigid if/else chain can’t express cleanly.
Frequently Asked Questions
Do I need to connect a tool for the AI Agent node to work?
Yes. n8n’s own documentation states you must connect at least one tool sub-node to an AI Agent node — without one, it functions as a plain chat node rather than an agent that can take action.
Does n8n AI agent memory persist after I close the workflow?
Not with Simple Memory, which n8n’s docs describe as using the running n8n instance’s own memory and not persisting between separate sessions. For memory that survives restarts, a database-backed memory option is the more durable choice.
What happened to n8n’s Conversational Agent node?
It’s deprecated as of n8n 1.82.0 — all AI Agent nodes now use the Tools Agent instead. Only workflows still running version 1 of the AI Agent node can still select it, and that version is removed entirely in n8n 3.0.
Can I stop an n8n AI agent from taking risky actions automatically?
Yes. For any tool connected to the AI Agent that performs a sensitive action, n8n lets you require human approval before that specific tool executes, configurable directly on the tool’s connector.
Can one n8n AI agent call another agent as a tool?
Yes, using the AI Agent Tool node or the Call n8n Workflow Tool, one agent can delegate a sub-task to another agent or workflow and use its output as part of its own reasoning — useful for splitting a complex task into smaller, more reliable specialist agents rather than asking one agent to handle everything.
Building a working n8n AI agent comes down to three connections done correctly — a chat model to think, memory to remember, and at least one tool to act — plus knowing when a reasoning agent is actually the right tool versus a standard, predictable workflow.
Sources
- n8n AI Agent node — official docs
- n8n official AI workflow tutorial
- n8n Tools Agent — official docs
- n8n Conversational Agent (deprecated) — official docs
- n8n 3.0 breaking changes — official changelog
- n8n Postgres Chat Memory — official docs
- n8n Redis Chat Memory — official docs
- n8n AI Agent Tool node — official docs
- n8n MCP Client Tool node — official docs
