Flow logic
Branch on a condition#
Two nodes split a flow, and picking the wrong one produces workflows nobody can read later.
If takes one condition and gives you two outputs, true and false. Use it when the question is genuinely binary.
Switch takes one value and routes to many outputs. Use it when you are dispatching on a category — an order status, a country, an event type. Four chained If nodes is a Switch that has not been noticed yet.
// If node condition, expression mode
{{ $json.status === 'active' && $json.credits > 0 }}An output with nothing connected to it ends that branch. That is a legitimate way to drop items you do not want.
Loop over items#
Most of the time you do not need a loop. When a node receives 200 items it runs 200 times on its own — iteration is the default, not something you request.
You need Loop Over Items (Split in Batches) when:
- An API rate-limits you and you must send 10 at a time rather than 200 at once.
- Each iteration depends on the result of the previous one.
- Memory matters and you would rather not hold the whole set at once.
Set Batch Size, connect the loop output back into the node, and take the done output when the set is exhausted.
Merge parallel paths#
Merge recombines two inputs. The mode decides what "recombine" means, and this is where most confusion lives:
| Mode | What it does | Use it when |
|---|---|---|
| Append | Stacks input 2 after input 1 | You want one combined list |
| Combine by matching fields | Joins on a key, like a SQL join | Each side holds part of the same record |
| Combine by position | Pairs item 1 with item 1, and so on | Two lists in a known, identical order |
| Choose branch | Passes one input, waits for both | You need synchronisation, not data |
Combine by position is the one that quietly breaks: it assumes an ordering guarantee that APIs rarely make. Match on a field whenever a field exists.
Wait#
Wait pauses execution — for a fixed interval, until a specific time, or until a webhook is called.
The third variant is what makes human approval possible: the workflow parks, someone clicks a link in the message you sent them, the webhook fires, and the run continues from exactly where it stopped. That pattern is covered in AI agents and chains, where letting a model act unsupervised is often the wrong call.
Split work into sub-workflows#
When a canvas stops fitting on a screen, extract part of it into its own workflow and call it with Execute Sub-workflow. You get reuse and a smaller surface to reason about.
One billing consequence to keep in mind: a sub-workflow run counts as its own execution. Splitting a workflow in three means three executions per logical run.
Related#
- Work with data — reshape what flows between these nodes
- Handle errors — what happens when a branch fails
- Core concepts — items and executions