n8n vs AI agents: when to use workflows, OpenClaw or Hermes

Use n8n when you can draw the path on a whiteboard before it runs. Use an agent like OpenClaw or Hermes when you can't, because the input is messy or the next step depends on what the last one turned up. Most people who run either for a few months end up with both: the agent handles the conversation and the judgement calls, and n8n does the writes that have to happen the same way every time.

The difference, in one sentence each

A workflow is a path you design: a trigger, a fixed set of nodes, an end. Feed it the same order twice and it does the same five things twice. That's why you can test it and predict what it costs.

An agent is a loop: a language model reads a goal, picks a tool, looks at what came back, decides what to do next, and keeps going until it thinks it's done. In n8n's docs, a chain follows a predetermined sequence while an agent "uses a language model to determine which actions to take." That loop is what lets an agent cope with an email that mixes a refund request with a complaint and a change of address. It's also why two runs on the same input can take different routes.

Neither one is smarter. They fail in different ways, and that should drive the choice more than any feature list.

What each tool actually is in November 2026

n8n

n8n is a workflow automation platform you can self-host or rent as n8n Cloud. It's fair-code rather than open source: the code is public, but the Sustainable Use License lets you use and modify it "only for your own internal business purposes or for non-commercial or personal use." Running it for your own business is fine. The current major version is 2.x; 2.45.7 was the latest stable release at the end of October 2026.

The core is the canvas: triggers such as Webhook and Schedule Trigger, hundreds of built-in nodes for specific apps, and an HTTP Request node for anything without one. The AI parts sit on top:

• The AI Agent node. Give it a chat model, at least one tool and optionally memory, and it decides which tools to call. Since version 1.82 every AI Agent node runs as a "Tools Agent".
• Single-purpose LLM nodes (Basic LLM Chain, Text Classifier, Information Extractor) for when you need one AI step, not a loop.
• MCP both ways. The MCP Client Tool lets an n8n agent use outside MCP servers, and the MCP Server Trigger exposes your workflows to outside MCP clients.
• A Chat Trigger, a Guardrails node, evaluations against test cases, and per-tool human approval for agents.

One newer feature blurs the line. Since 2.32.3, n8n also has standalone Agents with their own instructions, tools, skills and memory, reachable through Slack, Telegram, Linear or Microsoft Teams. As of November 2026 they're labelled Preview, and the docs say queue mode isn't supported for agents yet. Worth a try if you already live in n8n.

OpenClaw

OpenClaw is an MIT-licensed personal assistant that runs as one Gateway process and answers you in the chat apps you already use; its README lists Discord, iMessage, Slack, Teams, Telegram, WhatsApp "and 20+ more". It has real tools (exec for shell commands, a browser, web search and fetch, file edits) and skills: SKILL.md instruction packs that teach it a repeatable procedure.

Older guides mention the heartbeat, a periodic check-in that let the agent act unprompted. That's been retired. Heartbeat monitors are now ordinary Automations: scheduled jobs (cron expressions or one-shot times), inbound webhooks and Gmail triggers, each with its own delivery settings and run history.

Hermes Agent

Hermes Agent is Nous Research's MIT-licensed agent, and its big idea is the learning loop. When it works out a multi-step procedure, finds a way past a dead end, or gets corrected by you, its instructions tell it to save that as a skill with the skill_manage tool, and it patches those skills as it learns more. skills.write_approval: true makes each skill write wait for your review; turn it on early.

Memory is small on purpose: a MEMORY.md capped at 2,200 characters for the agent's own notes, a USER.md capped at 1,375 characters for what it knows about you, and a searchable SQLite store for older conversations. The messaging gateway covers more than twenty platforms, from Telegram and Discord to Signal and email. Its cron scheduler runs each job in a fresh session and delivers the result to whichever platform you name.

If you're choosing between the two agents, our OpenClaw vs Hermes comparison goes deeper.

Side by side

Six scenarios, six verdicts

1. Copying new orders into a spreadsheet

Verdict: n8n. There's nothing to decide. A trigger fires, you map fields, you append a row. An agent would do it slower, pay tokens on every order, and one day decide that a column called "Total" probably meant the tax-inclusive figure. Build it once, test it with pinned sample data, attach an error workflow, move on.

2. Triaging a messy support inbox

Verdict: n8n with an AI step, until the replies need judgement. "Put this email in one of six queues" is one decision with a fixed set of answers. A Text Classifier node in a workflow handles it. It stops fitting when the right action depends on context: the customer's last three orders, a refund they already got, a reply in your team's tone. That's agent work, ideally with an approval step before anything is sent.

3. A Telegram assistant that remembers context

Verdict: OpenClaw or Hermes. This is what they're built for: memory across conversations, Telegram and other chat apps out of the box, scheduled jobs that report back into the same chat. You can build a Telegram bot in n8n from a Telegram Trigger, an AI Agent node and a memory sub-node, but you'd be rebuilding per-user sessions, sender pairing and long-term memory that the agents already ship with.

4. A nightly sales report

Verdict: n8n. A Schedule Trigger, a couple of queries, a formatted message to Slack or email. It runs at 6:00, costs nothing extra, and when a number looks wrong you open last night's execution and see exactly what each node returned. Want a paragraph of commentary? Add one LLM node at the end to summarise the figures, and keep the arithmetic out of the model. Agents can run scheduled jobs too, but for numbers someone will act on, use the workflow.

5. Research with unpredictable steps

Verdict: an agent. "Find the most widely used open-source vector databases, check which shipped a release this quarter and summarise the breaking changes" can't be drawn in advance, because you don't know which pages matter until you've read some. Both agents have web search and browser tools, so they can follow a lead. Put a token budget or step limit on the job and read the sources it cites. Agents are very good at sounding finished.

6. Approvals with an audit trail

Verdict: n8n. Purchase requests, refunds over a threshold, publishing a page: anything where someone will later ask who approved it and when. The Slack node's Send and Wait for Response operation pauses the workflow until an approver clicks, and each execution keeps every node's input and output for as long as your retention settings allow. You can show an auditor one specific run. Explaining why an agent did what it did is a much harder meeting.

Using them together

The split that works: the agent owns the conversation and the judgement, and n8n owns anything that has to happen exactly once and exactly right. The agent decides a lead is qualified. n8n creates the contact, checks for duplicates and posts to the sales channel.

Option 1: the agent calls an n8n webhook

In n8n, start the workflow with a Webhook node set to POST, with Header Auth so only your agent can call it. Set Respond to "Using 'Respond to Webhook' Node" and return a small JSON result, so the agent gets a clear success or error instead of "Workflow got started." Once you publish the workflow, it's served on the production URL, by default https://your-n8n-host/webhook/<path>.

On the agent side, a skill tells the model when to call it and how. In OpenClaw that's a SKILL.md file in the workspace skills/ folder:

---
name: log-lead
description: Save a qualified sales lead to the CRM through the n8n lead workflow.
---

When the user confirms a lead, use the exec tool to run:

curl -sS -X POST "https://n8n.example.com/webhook/new-lead" \
  -H "X-Agent-Token: $N8N_LEAD_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"name": "...", "email": "...", "notes": "..."}'

Tell the user the "status" field from the JSON reply. If it is not "ok",
report the error and do not retry without asking.

Keep the token out of the file; OpenClaw can inject it for that one skill through skills.entries.<name>.env. Hermes skills are SKILL.md files with YAML frontmatter too, and Hermes runs the same command through its terminal tool.

Option 2: expose workflows over MCP

If you'd rather the workflow show up as a named tool, put an MCP Server Trigger in front of it. It publishes the tools attached to it, whole workflows included, on an MCP URL protected by bearer or header auth. Both agents connect to remote MCP servers: OpenClaw under mcp.servers in its config, Hermes in ~/.hermes/config.yaml:

mcp_servers:
  n8n:
    # paste the production URL shown on the MCP Server Trigger node
    url: "https://n8n.example.com/your-mcp-path"
    headers:
      Authorization: "Bearer YOUR_TOKEN"
    tools:
      include: ["create_lead"]

The include list matters more than it looks, because an agent can't misuse a tool it can't see. Hermes exposes the tool as mcp_n8n_create_lead so it never collides with built-in names (Hermes MCP docs).

The other direction

n8n can hand work to an agent too. OpenClaw's Gateway has inbound webhooks, off by default, that let an outside service submit an agent turn to /hooks/agent with a dedicated bearer token. That's handy when a workflow meets something its rules can't sort. For single AI steps you don't need a separate agent: a Basic LLM Chain or AI Agent node calls the model inside the workflow.

Cost and reliability

What you pay

Self-hosted n8n Community edition is free with no per-execution fee; n8n's docs call it "almost the complete feature set." SSO, environments or Git version control need a paid Business or Enterprise licence. You pay for the server and do the patching.

n8n Cloud charges by executions, and one execution is one run of a whole workflow however many nodes it has. As of November 2026 the pricing page lists Starter at €24 a month (€20 billed annually) for 2,500 executions and Pro from €60 a month (€50 billed annually) for 10,000. Business, €667 a month billed annually, is self-hosted only. Turns with n8n's own Agents count against the same quota.

An agent's bill has two parts. Hosting is flat; tokens aren't. Every turn sends the instructions, relevant memory, the conversation so far and the tool definitions, and one request can take several model calls as the agent works through its tools. Our breakdown of agent LLM costs per month has worked numbers. An AI step inside n8n costs tokens too, but it's bounded: one call per item.

How each one fails, and the guardrails that help

Workflows fail loudly. An API renames a field, a credential expires, a payload arrives in a new shape, and the run stops at a red node. Annoying, but honest. What helps:

• Attach an error workflow (an Error Trigger that posts to Slack or email) to every production workflow.
• Check the payload at the top with an IF node and reject what you don't recognise, before a null travels three nodes downstream.
• Turn on retries for flaky API nodes, and keep pinned sample data so you can re-test after a change.

Agents fail quietly, sometimes creatively. The model picks the wrong tool, misreads a request, retries something that shouldn't be retried, or follows instructions hidden in a page or email it was asked to read. What helps:

• Give each agent the fewest tools it needs, and filter MCP tools with an include list.
• Require approval for anything irreversible. OpenClaw has exec approvals. Hermes has approvals.mode, smart by default or manual if you want a prompt every time. n8n's AI Agent node can require human review per tool.
• Route writes through an n8n webhook that validates input and enforces limits, so a confused agent gets an error instead of writing a bad record.
• Decide who can message it: pairing or allowlists on every chat channel.

A quick decision guide

If you can draw it, build it in n8n. If one step is fuzzy, add one AI node. If the whole task is fuzzy, use an agent, and let n8n do any write that matters.

When you need only one of them

The overlap is real. A handful of integrations plus the occasional AI summary is an n8n-only job. A Telegram assistant that knows your schedule is an agent-only job. The combined setup earns its keep once the assistant starts doing things with consequences.

That combined setup is a normal one on HostAiHub. We host all three: n8n from $7 a month, and OpenClaw and Hermes at $10 a month each. Each runs in its own private container on its own SSL subdomain and deploys automatically in a few minutes, so an agent next to an n8n instance is just two services talking over HTTPS. n8n plans include unlimited workflows and executions on your instance, plus daily backups. The agents use your own LLM key, or our OpenAI-compatible AI API.