Running
Test runs that change nothing, what each step did, and the limits.
A test run
Test an automation before you let it loose. A test run reads everything for real and writes nothing: no rows are added, no email is sent, no webhook is called, and a folder of documents is left untouched so each one is still waiting for the real run.
It also leaves the clock alone. A test does not count as the night's run, and it does not move any watermark, so the first real run still reads everything it should.
What each step did
Every run keeps a trace, step by step: what arrived, what left, and a sample of the rows. For a step that acted on the outside world it records what was sent, with the host named rather than the full address.
The sample is a slice rather than everything, and it says when it is a slice. There are two caps, on rows and on size, and whichever bites first wins, so a very wide table is cut even when it has few rows.
For an AI step, the trace holds the prompt that was actually sent. A test run's trace says plainly where it substituted a shorter one, because a prompt that has been cut without saying so is worse than none: you read it, conclude the instructions are fine, and go looking elsewhere.
Run history
Each automation keeps its runs, newest first, with how long each took, what it cost, and which step failed if one did. A run started by a schedule records nobody as the actor, because nobody was watching; one you started records you.
A failing step names the cause where it can, including the case that used to be hardest to read: a step with nothing to work on now names the source above it that has no table chosen, rather than blaming itself.
Limits
| Free | Paid | Business | |
|---|---|---|---|
| Automations | 3 | Unlimited | Unlimited |
| Runs at once | 1 | 10 | 20 |
Downgrading keeps the automations you already have. Only new ones are blocked past the cap.
There is also a monthly ceiling on runs, per seat and pooled across a business, which exists as a backstop against an automation that somehow loops rather than as something ordinary use reaches.
What a run costs
Runs are billed in credits to the space that owns the automation, and appear in your usage history as automation runs. Test runs are not counted against your allowance.
The cost is what the steps do. Reading and reshaping tables is close to nothing. An AI step is where the cost is, and one set to run per item is a call for every row or document, which is why that mode is capped.
You can set a ceiling on what one automation may spend. Past it, runs stop rather than continue quietly.
When something is wrong
A few things are worth knowing about, because each is quiet by nature:
- A trigger pointed at nothing never fires, and looks finished. The canvas marks the step, and the automation's status says it is not pointed at anything yet.
- A source with no table chosen fails on the first run, and the message names it.
- A schedule below your plan's floor is slowed on save, with a note.
- A step that sends keeps a watermark, so the same rows are not sent twice. A test run never moves it.