How an assistant makes live changes safely

Assistants only edit drafts

An assistant can build and edit Draft and QA experiments, and Draft surveys. It cannot change an experiment that is running, deployed, ended or archived, or a survey that is running or ended. To change one, it duplicates it to a new draft, or asks a person to make the change.

That means nothing an assistant edits is ever sitting on a live site, waiting for someone's next publish. You can still edit any experiment yourself, as usual.

Preview, then confirm

This applies to starting, ending, scheduling, deploying and archiving experiments, starting and ending surveys, and publishing a website.

1. 1The assistant asks Pertento to do it. Nothing happens yet: Pertento answers with a preview in plain words.
2. 2The assistant is asked to show you the preview and its warnings.
3. 3To go ahead, it confirms with a second request. Pertento checks that nothing has changed since the preview, and only then acts.

A confirmation lasts 5 minutes and works once. If anything changed in between, for example someone edited a weight, it is refused with "Changed since the preview", and the assistant has to preview again.

Pertento does not ask you a second time itself. Whether your assistant asks you before each request is set in the assistant's app, and most let you choose. If you want to approve every live change, set that there, or leave the scopes that make live changes out when you connect.

What the preview shows

• Starting an experiment: the traffic split, where and on which devices it shows, and sessions left this month. It warns when the account has used its sessions, so the experiment would start but show to nobody, and when a variant looks the same as the original.
• Ending: how many days it has run, and a warning that ending a deployed experiment takes the winner off your site.
• Scheduling: the times, the split, and any schedule it replaces.
• Deploying: the verdict and the recommended winner. It warns when the chosen variant is not the recommended one, is not significant, is significantly worse, or is the original.
• Archiving: a warning that archiving does not stop a running or deployed experiment.
• Starting a survey: its questions, where it shows, the sample rate and the sessions left.
• Publishing: what will be live afterwards, and what stays in QA.

A publish ships everything

Starting, ending, deploying, switching QA on or off, starting or ending a survey, and publishing all publish the website, as they do from the dashboard. A publish ships everything set on the website at that moment, including an edit a person made to a running experiment and has not published yet.

You are told when your assistant acts

When an assistant starts, ends or deploys something, you get a notification that says your agent did it, with the token's name. Colleagues see that it was your agent. In the experiment's activity log, the line is marked "via agent".