Building AI Agents and Sub-Agents in n8n (2025 Guide)

The Rise of Agentic Automation

Traditional automation is deterministic: if X, do Y. AI agents change the game — they can reason about a task, decide which tools to use, handle unexpected inputs, and loop until they've achieved a goal.

n8n has first-class support for AI agents through its LangChain integration. You can build agents that use GPT-4, Claude, or any other LLM as the reasoning engine, with n8n tools available as capabilities.

Understanding n8n AI Agents

An n8n AI Agent consists of:

A language model (OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint)
Tools — n8n nodes the agent can call (HTTP requests, database queries, email, etc.)
Memory — optional conversation history (in-memory, Redis, or external)
A system prompt — instructions that define the agent's role and behaviour

Building Your First Agent

Step 1: Add the AI Agent Node

Drag in the "AI Agent" node. Connect it to an LLM node (e.g., OpenAI Chat Model with gpt-4o) and optionally a memory node.

Step 2: Define Tools

Tools are what the agent can "do". In n8n, you can use:

Call n8n Workflow tool — calls another n8n workflow as a tool
HTTP Request tool — makes API calls the agent decides to make
Calculator, Wikipedia, SerpAPI — built-in tools
Custom tools — any n8n node wrapped as a tool

Step 3: Write the System Prompt

Be specific about what the agent should and shouldn't do. Good agents have tight system prompts:

You are a customer support agent for n8nautomation.cloud.
Your job is to answer questions about n8n, check subscription status,
and escalate complex issues to the human support team.
Always be concise. Never make up pricing information.

Sub-Agent Patterns

The real power comes from multi-agent architectures. A orchestrator agent delegates to specialised sub-agents:

Orchestrator → Specialist Pattern

The orchestrator receives a high-level task ("research competitors and write a report"), breaks it into sub-tasks, delegates them to specialist agents (a researcher agent, a writer agent), and assembles the final output.

In n8n, implement this using the "Call n8n Workflow" tool — each sub-agent is a separate workflow that the orchestrator calls with specific instructions.

Parallel Sub-Agents

For tasks that can be parallelised (e.g., processing multiple customer queries simultaneously), trigger multiple sub-agent workflows in parallel using n8n's branching, then merge results.

Reviewer Pattern

Have a second agent review and critique the first agent's output. This dramatically improves quality for writing, code generation, or analysis tasks. The reviewer agent returns structured feedback, and the generator agent revises its output.

Memory Strategies

Window Buffer Memory: Keeps the last N messages. Good for short conversations.
Redis Memory: Persists across executions. Essential for long-running agent sessions.
Vector Store Memory: Semantic search over conversation history. Useful for agents that need to recall facts from many turns ago.

Production Considerations

AI agents are non-deterministic — they can take unexpected paths. For production deployments:

Set maximum iteration limits to prevent infinite loops
Log every agent step to a database for debugging
Add human-in-the-loop steps for high-stakes decisions (sending emails, making purchases)
Use structured output parsing to ensure the agent returns data in the expected format
Monitor token usage — agents can burn through LLM API budgets quickly

Run Agents on a Dedicated Instance

AI agent workflows often run for minutes rather than seconds. A shared n8n instance has execution timeouts that will interrupt long-running agents. A dedicated n8n instance lets you configure execution timeouts to match your use case — no shared resource constraints, no interruptions mid-reasoning.

Building AI Agents and Sub-Agents in n8n: A Practical Guide