Automation workflows start small. Three months later you have forty nodes, the same phone-number-cleaning logic copied to four places, and nobody wants to touch it. The solution mirrors programming: extract functions. In n8n, these functions are sub-workflows.
The real question is not whether to decompose, but where to draw the boundary. Every split creates an interface contract. Get it wrong and you trade spaghetti nodes for spaghetti state-passing. Get it right and you have reusable components that double as agent tools.
When to Split: Signals and Anti-Patterns
Split when you see:
The same node sequence in multiple workflows
A logical unit you can name in one sentence ("upsert contact in CRM")
A step you need to test independently
An action an AI agent should invoke on demand
Do not split:
Two-node sequences used once (pure overhead)
Tightly coupled state that requires the parent's full context
Logic that changes every time the parent changes
The boundary test: can you write a clear input…