n8n's AI Agent Builder (2.40 Preview): First Look + Real Cost Math

Some links in this article are affiliate links. Automated Wealth Lab may earn a commission at no extra cost to you. See our Affiliate Disclosure.

Contents

What’s actually in the preview

Old pattern vs. new artifact
Why this matters for solo founders

The GTC Berlin hook: announced, not attended
The real cost math

The platform layer
The model layer (what neither vendor includes)
Worked example: 1,000 support requests a month

The approval pattern: why the NemoClaw demo matters
FAQ
Methodology
AWL verdict
Sources

Disclosure: Automated Wealth Lab is reader-supported. n8n is one of our affiliate partners — if you start n8n Cloud through n8n, we may earn a commission (n8n pays 30% of Cloud referrals for the first year, per their public affiliate page) at no extra cost to you. This first look is documented from public release notes and announcements — no hands-on testing, and we say so.

n8n has been an automation platform for a decade: triggers, workflows, deterministic steps. On September 25, 2026, it added a second surface: Agents — standalone AI agents that sit beside workflows instead of inside one particular canvas. The timing matters: n8n is also scheduled to appear at NVIDIA GTC Berlin (October 20–22, 2026) demonstrating governed agent actions with NVIDIA’s NemoClaw.

This is a documented first look, not a review. Nothing below was hands-on tested. Every claim traces to release notes, docs, or announcements; the cost math comes from published price sheets with the assumptions labeled.

What’s actually in the preview

n8n’s boundary line is the most useful sentence in the launch: the new Agents surface is for work that is too open-ended for a fixed workflow — inbox review, ticket triage, a Telegram conversation that needs follow-up questions. Jobs where the path depends on what’s found along the way.

Old pattern vs. new artifact

The old pattern still works: a Chat Trigger, memory, an AI Agent node, a model, and tools assembled inside a workflow canvas.

The preview changes ownership. An Agent is now a first-class artifact with its own model and instructions, tools, memory and sessions, draft and published versions (editing the draft doesn’t silently change the published agent), channels (Slack, Telegram, Discord, Linear), schedules — and the ability to be invoked from another workflow. One agent can answer in a channel, run on a schedule, or take a workflow message; sessions log the conversation, tool calls, errors, and pending approvals.

n8n 2.40 (September 15, 2026) added the manual agent-creation entry point; n8n 2.42 (September 29) refined the builder — layouts, sub-agent guidance, activity feedback, compact configuration.

Why this matters for solo founders

The most commercially interesting line in early reactions: reusing existing workflows as tools beats adding another chat wrapper. An agent with a narrow tool contract — a workflow that looks up an account, drafts a reply, logs a record — inherits your existing automation instead of replacing it.

The documented design rule: don’t give the agent broad access to every system. The agent owns the conversation and tool-selection loop; a workflow owns the bounded action; an approver owns the final decision for anything sensitive. That separation is the whole game.

The GTC Berlin hook: announced, not attended

Flat-lay: frosted glass jar with a padlock charm holding a lavender paper crane beside a coral approval stamp, symbolizing a sandboxed agent awaiting human approval

Precision matters with event news. NVIDIA GTC Berlin runs October 20–22, 2026. Two n8n items are on the program:

Session	When	What it is
“Bring Your Own Agent: Sovereign Guardrails for the Age of Autonomous Claws”	Tue Oct 20, 11:00am–1:00pm CEST	Talk by n8n Director of Engineering Claire Knight on sandboxing and governance for agent workflows
“Your Agent Doesn’t Get Root: Governing AI Actions With n8n and NVIDIA NemoClaw”	Thu Oct 22, 3:00–4:45pm CEST	Hands-on lab: a sandboxed agent whose actions are exposed through n8n workflows and require human approval

Confirmed: n8n is publicly positioning itself as the governance layer for AI agents. The lab runs the agent in a locked-down NemoClaw/OpenShell sandbox with no direct network or API access — actions go through n8n workflows, an operator approves them, and the model records who approved what and where the request originated.

Not confirmed: no new product release, pricing change, or GA announcement — a demonstration of an approach, written before the event.

The real cost math

Agent pricing is where vendors hide the second meter (prices checked October 2026).

The platform layer

Platform	Plan	Price	Monthly allowance
n8n Cloud	Starter (annual)	€20/mo (~$22–23)	2,500 executions
n8n Cloud	Pro (annual)	€50/mo (~$56–57)	10,000 executions
n8n Community	Self-hosted	$0 (server ~$5–10/mo)	Unlimited (server-bound)
Zapier Agents	Free	$0	400 activities
Zapier Agents	Pro (annual)	$33.33/mo ($400/yr)	1,500 activities

Three meter differences that decide the math:

n8n counts executions, not steps. One run = one execution, however many nodes. Zapier counts tasks per action step.
One Zapier agent run can consume several activities — browsing, lookups, actions each meter.
n8n has no separate AI tier. Agent runs consume executions like everything else. Zapier Agents is a separate add-on with its own allowance, independent of your Zap task pool.

The model layer (what neither vendor includes)

Both let you bring your own model key: the vendor bills orchestration, the model provider bills tokens. Any “X costs $Y/mo” claim that ignores inference is incomplete. The honest budget is platform fee + measured token spend — and token spend is unknowable until you run.

Worked example: 1,000 support requests a month

Assumptions, stated plainly — illustrative, not measured:

Each request = 1 n8n agent execution
Each request = ~5 Zapier activities (lookup, knowledge check, draft, send, overhead)
Inference excluded from both — roughly equal either way with your own provider key

	n8n Cloud Starter	Zapier Agents Pro
Platform cost	€20/mo (~$22–23)	$33.33/mo
Allowance	2,500 executions	1,500 activities
Requests covered	1,000 (40% of allowance)	~300 (1,500 ÷ 5)

At this volume, n8n Starter covers the workload at roughly two-thirds the price; Zapier Agents Pro caps out around 300 requests. The free tier’s 400 activities covers roughly 80.

Two caveats: chatty agent runs worsen the activity math, and n8n agents fanning out into many sub-executions shift the execution math. Self-hosted n8n Community makes the platform line $0 plus a $5–10 server — then the only variable cost is inference. And confirm which Zapier agent product you’re evaluating: standalone Agents are migrating into the Zap editor under “AI by Zapier” with model-tiered task pricing — a different meter.

The approval pattern: why the NemoClaw demo matters

The lab is a reusable pattern for any agent platform:

Sandbox the agent. No direct network or API access. In NemoClaw that’s OpenShell isolation; in your n8n setup it’s never handing the agent a broad-scope credential.
Expose actions as narrow workflows. Named tools with defined inputs, controlled credentials, known outputs. The agent can only do what a workflow lets it do.
Put a human on the sensitive ones. The agent requests; the operator approves; the system records. n8n’s sessions log pending approvals alongside the conversation, so the audit trail exists by construction.
Record provenance: who approved, which agent, which input. When something breaks at 2am, that’s the difference between a fix and a forensics project.

This is the unsexy half of the agent hype cycle, and it’s the half that determines whether you can deploy an agent on customer-facing work. A solo founder can’t afford an agent that refunds the wrong customer. Approval boundaries are what make agents usable, not just impressive.

FAQ

Is the Agent Builder generally available?

No — preview on n8n Cloud (2.40/2.42, September 2026). Expect rough edges and changing behavior; don’t build revenue-critical workflows on it yet.

Do agents cost extra on n8n?

No separate agent tier exists. Agent runs consume executions on whatever Cloud plan or self-hosted setup you have. Inference is billed by your model provider, not n8n.

Is this better than Zapier Agents?

Different meters, different philosophies: n8n runs executions with no AI surcharge and self-hosting; Zapier bills activities on a separate $33.33/mo add-on with a bigger catalog. For real agent volume, n8n’s math is usually kinder.

Is the GTC Berlin demo a product launch?

No evidence of that. The confirmed items are a talk and a hands-on lab on governed agent actions. Treat it as n8n staking a position on agent governance, not a release.

Can I use the Agent Builder self-hosted?

The preview is documented for n8n Cloud. Self-hosted availability of preview features typically lags — check the release notes for your version.

Does the Agent Builder replace n8n workflows?

No — per n8n’s framing, agents handle open-ended work while workflows keep deterministic work. The power move is agents calling workflows as tools.

Methodology

A documented first look — no hands-on testing; we had no preview access.

Verified (October 2026): n8n 2.40 (2026-09-15) and 2.42 (2026-09-29) release dates via GitHub release tags; Agents surface launch (September 25, 2026) and the first-class-artifact model via n8n’s Agents documentation; channels (Slack, Telegram, Discord, Linear) and entry points via the official announcement; GTC Berlin sessions, dates, times (framed strictly as announced/upcoming); NemoClaw sandbox model (no direct network/API access, actions via n8n workflows, human approval, approval provenance) via the lab description; n8n Cloud prices (Starter €20/mo annual / 2,500 executions; Pro €50/mo / 10,000; executions = full workflow runs) via multiple independent Oct 2026 price checks; Zapier Agents pricing (Free 400 activities; Pro $33.33/mo annual / 1,500 activities) via September 2026 documentation; n8n affiliate terms (30% of Cloud referrals, first year) via n8n’s public affiliate page — marker [AFFILIATE LINK: n8n] inline, owner adds the URL.

Not verified: Agent Builder behavior, reliability, latency, UX — not tested. Per-request activity/execution consumption — the worked example uses labeled assumptions, not measurements. Inference costs — provider-dependent, excluded from platform comparisons.

AWL verdict

AWL verdict: The Agent Builder preview is the most interesting thing n8n has shipped in years — not because agents are magic, but because the architecture is right: agents as governed artifacts with draft/published versions, narrow workflow tools, and approval boundaries, all on the same execution meter. The cost math favors n8n over Zapier Agents for real volume, especially self-hosted. But it’s a preview: prototype with it, don’t bet the business on it until GA. The October 22 GTC Berlin lab is the governance story’s first public stress test.