Testing and publishing
Every agent has two versions: the draft you are editing and the published snapshot that actually runs. Understanding the split is what lets you edit a live agent without fear.
Preview mode — nothing is sent
The Preview button (eye icon) runs the draft in dry-run mode. Every tool call that would write something — send an email, post to Slack, send a WhatsApp — is simulated instead. The transcript shows what would have happened, with each simulated action labelled in amber. Read tools (reading mail, reading a spreadsheet) still execute normally so you see realistic data.
Use this to check new instructions before publishing. No accounts are touched and nothing costs money.
Live runs are real
The Run button is live — emails are sent, messages go out, texts cost money. The only thing it changes compared to a scheduled run is which version runs: the draft rather than the published snapshot.
Draft and published
- The draft is what you see in the builder. Every edit saves here immediately.
- The published snapshot is a frozen copy taken at the moment you pressed Publish. Scheduled runs use this and nothing else.
So you can rewrite the instructions of a live agent in the middle of the day, and the version running on the schedule stays exactly as it was until you decide to publish. When the two differ, the agent shows an unpublished-changes marker.
Which version runs when
- Test run - the draft. This is the point: you are checking the edit you just made.
- Scheduled run - the published snapshot, always.
Publishing
Two things must be true before an agent can be published:
- An AI model is selected.
- The instructions are not empty.
Publishing takes the snapshot and marks the agent live. Its schedule starts firing from the next occurrence. Note that connectors are not part of the snapshot - attaching or detaching an account takes effect on the very next run, published or not.
Pausing and resuming
Pausing is a softer stop than unpublishing. A paused agent keeps its published snapshot intact — it just skips every scheduled and message-triggered run until you resume it. The amber Paused badge appears on the card and in the header.
- Pause — available on any live (published) agent. The agent keeps its configuration; nothing is lost.
- Resume — returns the agent to live immediately, using the same published snapshot. No new publish step required.
- Unpublish — clears the published snapshot entirely. You will need to publish again before the agent can run on its own.
Use Pause when you need the agent to go quiet for a day or a week and you plan to resume it unchanged. Use Unpublish when you want to rework the agent properly before it runs again.
Unpublishing
Unpublishing takes the agent off its schedule and puts it back into draft. Nothing is lost - the instructions, tools and triggers all stay, they just stop running by themselves. This is the right move when an agent is misbehaving and you need it to stop now.
A sensible first-run routine
- Set the daily budget low while you are still experimenting.
- Write instructions that report rather than act - “tell me what you would do” - and test those first.
- Read the transcript. Check it looked at the right things.
- Change the instructions to actually act, and test again.
- Publish, then watch the first two or three scheduled runs.