n8n vs. LangGraph — which layer owns the AI workflow

Same job title, completely different design targets

n8n and LangGraph both build automations, and the similarity ends there. n8n is the visual automation layer: an editor where you drag and drop steps — a webhook trigger, an API call, a filter, an AI agent node — and connect them without writing code. LangGraph is the code-first agent framework: you compose a graph in Python where LLM nodes decide the routing, tools execute on the model's behalf, and cycles repeat until a termination condition is met.

The comparison that helps is not which is better — it is which job each one actually does. The same team usually needs both, and the gateway is the wall between them.

The common pattern is both: n8n for the deterministic pipeline that ingests the data, LangGraph for the agent that decides what to do with it.

How the layers compose

A production automation usually has a deterministic spine and an agentic decision point at the centre. The spine — triggered by an email, a webhook or a schedule — is n8n territory: fetch, transform, filter, route. The decision point — classify the intent, decide which system to query, compose a response — is LangGraph territory.

The two layers connect through a webhook node: n8n calls the LangGraph endpoint with the context, LangGraph returns the agent's decision, and n8n carries the result to the next step. That split is what makes the automation both fast to build and deep enough to handle real-world variation.

The two layers meet at a webhook

n8n triggers the workflow, LangGraph reasons about it, n8n delivers the result.

// n8n workflow: HTTP Webhook → LangGraph agent → Slack
1. Webhook — receive customer email attachment
2. LangGraph Agent node — call the LangGraph endpoint
   { "messages": [email_body], "context": { "customer_id": u } }
3. IF Agent returned "escalate" THEN Slack to support team
4. ELSE update CRM and send confirmation email

Three questions that decide the workflow

Each is answered by watching your team work.

• Is the workflow primarily deterministic or agentic? A fixed sequence of API calls is n8n territory; a sequence that changes based on LLM output is LangGraph territory.
• Who maintains it? Operations teams who prefer a visual editor point toward n8n; engineering teams who treat workflows as code point toward LangGraph.
• Does it need to resume after a human decision? If yes, LangGraph's state persistence is the cleaner model.

The most expensive mistake is picking one layer and forcing every use case through it. The stack is both.

Go deeper on the automation stack

• Open WebUI — the self-hosted chat interface — The front end that triggers the automations behind it.
• LiteLLM vs. OpenRouter — The gateway that logs every step of both layers.
• MLflow vs. Langfuse — The observability layer that tells you whether the automation is working.