n8n Agents can use your existing n8n workflows as tools, so an AI agent can look up NetSuite data and request NetSuite changes while every write still runs through a workflow you built and tested. The agent decides which tool to call; the workflow does the work, with the credential you attached to it; and any tool you mark as sensitive pauses for a person to approve or reject it first. n8n announced Agents on September 25, 2026, and the feature is still in Preview.
This guide covers what n8n Agents are, how they differ from the AI Agent node, how we would set one up for NetSuite, and the guardrails to put in place before it touches real records. Facts about n8n come from n8n's own blog and documentation (sources at the end). n8n's pages don't mention NetSuite, so the NetSuite design below is our recommendation, not n8n's.
What n8n Agents are
n8n describes Agents as sitting "next to your workflows". You build an agent in the Agent Builder: choose a model, write its instructions, and add tools. n8n gives it memory, sessions, channels, versions and approvals out of the box, which you would otherwise assemble yourself from the AI Agent node, a chat trigger, a memory node and tools.
n8n puts the choice like this: "Sometimes you want the workflow in charge, with the agent as a step inside it. Sometimes you want the agent in charge, with workflows as tools." Its docs say to use an agent "when work is too open-ended for a fixed workflow". The AI Agent node hasn't changed and isn't replaced.
| n8n Agents | AI Agent node | |
|---|---|---|
| Who is in charge | The agent, calling workflows as tools | The workflow, with the agent as one step |
| Memory | Session memory on by default; optional episodic memory across sessions (needs an OpenAI credential) | Whatever memory node you connect |
| Where people use it | Agent Builder chat, Slack, Telegram, Linear, Microsoft Teams, schedules, or the Message an Agent node in a workflow | Whatever trigger the workflow has |
| Approvals | Mark a tool as sensitive: the agent pauses for Approve or Reject | Human-in-the-loop review on all or selected tools (Approve or Deny), in Chat, Slack, Teams, Gmail and other channels |
| History | Sessions tab: messages, tools used, pending approvals; versions from each publish | Workflow executions |
| Status | Preview | Available today |
We covered the AI Agent node with NetSuite in n8n AI Agents for NetSuite: Beyond the Demo. This article is about the new, standalone Agents.
Why "workflows as tools" suits NetSuite
The risk with an AI agent and an ERP is not that the model can't call an API. It's that it calls the API in a way nobody tested. n8n's design addresses that directly: "Every workflow you've built in n8n is something an agent can use, and none of it needs to change." The agent picks when a workflow runs; the workflow runs as built.
For NetSuite that means three things:
- The write logic stays in a workflow. Field mapping, validation, SuiteQL lookups and error handling live in a workflow you can test with fixed inputs, the same way you test any other n8n NetSuite integration.
- The agent never holds the NetSuite credential. In n8n's words, "Each tool runs with the credential you attached to it. The agent never holds the keys to your instance." Access to the agent itself follows your n8n roles.
- Every write can wait for a person. Mark the tool as sensitive and the agent stops for Approve or Reject before it runs.
A NetSuite agent design: three tools
Here is how we would start a finance or customer-service agent for NetSuite. Each tool is its own workflow, does one job, and returns a short summary the agent can repeat back to the user.
| Tool (workflow) | What it does | Approval | Design notes |
|---|---|---|---|
| Look up customer | Returns a customer's open orders, unpaid invoices and credit limit | None (read-only) | Input: customer ID or email. Run fixed SuiteQL queries; don't let the agent write its own queries. |
| Draft credit memo | Creates a credit memo against one invoice, for an amount and reason | Required | Check the amount is no more than the invoice's remaining balance, and stop above a set limit. Return the new record's ID and link. |
| Update customer field | Changes one of a short list of allowed fields, such as email or terms | Required | Allow-list the fields in the workflow. Return the old and new value so the approver sees both. |
The pattern matters more than these three tools: narrow inputs, one record type per tool, limits enforced in the workflow rather than in the prompt, and output a person can check at a glance. A tool that "updates any NetSuite record" is the one to avoid, because it turns the agent's guess into an unreviewed write.
Setting it up
- Check availability. n8n's docs list Agents on n8n Cloud on all plans, and on self-hosted n8n on all plans from version 2.32.3. Self-hosted agents need a public, secure URL (the
N8N_WEBHOOK_URLsetting) before you can connect channels such as Slack. Agents don't support queue mode yet; n8n says to run them in regular mode for now. If you are planning the n8n 3.0 upgrade, test agents on the version you will actually run. - Build the tool workflows first. Put them in the same project as the agent, since the docs list "workflows in the same project" as a tool source. Test each one on its own with real NetSuite data in a sandbox account.
- Create the agent. Choose a provider and model you have a credential for, or use n8n's Gateway credits. Write instructions that say what the agent is for, which tool to use for what, and what it must not do.
- Add the tools and mark the writes as sensitive. Leave the read-only lookup without approval so the agent stays useful; require approval on every tool that changes NetSuite.
- Test in the Agent Builder chat, then publish. Publishing takes a snapshot of the draft, and every publish adds a version to the history. Schedules run only the published version.
- Connect a channel. Start with a test Slack or Teams channel. In Teams, an approval appears as a card that only the person who has to answer can see. (The Teams channel needs n8n 2.46.0 or later.)
Guardrails before it touches real records
n8n's own advice for the Preview is to "start where the blast radius is small: scoped tools, a test channel, and approvals on any action that writes to a system of record." For NetSuite we would add:
- A NetSuite role per agent. Give the credential behind the tool workflows a role that can do only what those tools do. If the agent has to be cut off, you can revoke that one integration without touching anything else; see what to do when an AI agent writes bad data to NetSuite.
- Approval on every write, including small ones. Remove approval from a tool only after it has a track record, and never from a tool that moves money.
- Two audit trails. The agent's Sessions tab shows the messages, the tools used and pending approvals. NetSuite's System Notes show what changed on each record. Review both during the first weeks.
- Limits in the workflow, not the prompt. A prompt that says "never credit more than $500" is a request. An IF node that stops the workflow above $500 is a control.
- Idempotency on writes. An agent can call the same tool twice. Check for an existing record before you create one, as covered in race conditions in n8n agent workflows.
For the wider policy side (who owns an agent, what it may do, how it is reviewed), see our AI agent governance guide.
What it costs to run
n8n counts one turn with an agent as one execution, and agents share the same execution quota as workflows. Tool calls to your workflows and to sub-agents don't count separately. Model usage is billed by your model provider, or drawn from Gateway credits if you use those. One thing to watch in Teams: with the option to read all messages turned on, "every message uses model credits, even when the agent doesn't reply." For plan quotas, see our n8n pricing breakdown.
n8n Agents, the AI Agent node or MCP?
- n8n Agents: people ask open-ended questions in chat, Slack or Teams, and the agent chooses among your tested NetSuite workflows. Best when the requests vary and the actions are well defined.
- AI Agent node: a fixed process with one judgement step, such as reading an incoming invoice and deciding which purchase order it matches. The workflow stays in charge.
- MCP: an outside AI client (Claude, ChatGPT or Cursor, for example) needs NetSuite access. See n8n as an MCP server for NetSuite. n8n Agents can also use tools from MCP servers.
Frequently asked questions
What are n8n Agents? n8n Agents are AI agents you build in n8n's Agent Builder. They come with memory, sessions, channels, versions and approvals, and they can use your existing n8n workflows as tools. n8n announced them on September 25, 2026, and they are in Preview.
How are n8n Agents different from the AI Agent node? With n8n Agents, the agent is in charge and calls workflows as tools. With the AI Agent node, the workflow is in charge and the agent is one step inside it. The AI Agent node hasn't changed.
Can an n8n Agent write to NetSuite? Yes, through a workflow tool that writes to NetSuite with the credential you attach to it. Mark that tool as sensitive so the agent pauses for a person to approve or reject each write.
Do n8n Agents work on self-hosted n8n? Yes. n8n's docs list self-hosted support on all plans from version 2.32.3. Channels need a public URL for the instance, and queue mode isn't supported for agents yet.
How much do n8n Agents cost? One turn with an agent counts as one n8n execution, from the same quota as workflows. Tool calls to workflows don't count separately. Model usage is paid to your model provider or through n8n's Gateway credits.
Sources
- Introducing n8n Agents (n8n blog, September 25, 2026)
- Build and manage agents (n8n docs)
- Microsoft Teams channel for agents (n8n docs)
- Message an Agent node (n8n docs)
- Human-in-the-loop for tools (n8n docs, AI Agent node)
All checked November 9, 2026. Agents are in Preview, so details can change; we'll update this page as they do.
Want a NetSuite agent built with these guardrails?
Entech Solutions has built production workflows on n8n for over two years, including NetSuite and AI-agent integrations. We build the tool workflows, set up the NetSuite role and approvals, and test the agent on your sandbox before it goes live. Talk to us or see our n8n services.