Assistants that work on their own

Some work starts before you ask: new mail that needs a reply, an issue opened overnight, a change to your calendar. With Works on its own, an assistant looks at the services you have connected on a schedule, does what you told it to do, and tells you when something needs you. You do not have to ask each time.

What it does#

• It checks on a schedule you choose, at most once an hour.
• Each check looks only at what is new in the services you picked since its last check.
• It does what your brief asks, such as reading and summarising, and suggesting replies, and writes what you should know as its answer.
• It tells you only when something needs you. A check with nothing new is one quiet line in its chat.
• Anything that sends, posts, spends, or publishes waits for you. A check never gets more freedom than a message you type yourself, and in one way less: it only reads, and a change in a connected service is turned back.
• Each check uses credits, like any other message.

Set it up#

You turn it on for one assistant at a time, in any of these places. Nothing runs until you press the button yourself.

• When you hire. The hire form has a Works on its own step after the team step. It is off until you turn it on.
• On a ready-made assistant. The one-click hire has an option to let the assistant work on its own. Ticked, the hire opens the assistant's page with the card ready, and nothing runs until you save.
• On the assistant's page. The Works on its own card sits under the assistant's details. Choose Set it up.
• From the chat. Ask the CEO assistant to keep an eye on something, such as "watch my inbox and tell me who needs an answer". It shows the same form as a panel in the chat, filled in for you. It saves nothing: you read it, change it, and press Save.

The form asks for three things.

You can set it up for up to 20 assistants.

Why only connected services#

A check can only read what you have connected. A service you have not connected is greyed out on the form, with a Connect link that opens Integrations in a new tab. Come back and the list updates by itself.

• A service can be connected and still unreadable for one assistant, because its role does not work with that kind of service yet. The form says so, and you can stop watching it or choose another assistant.
• GitHub connected with a token reads issues only. It does not read pull requests or checks.
• If a service you watch is disconnected later, the next check says it could not look. That counts as an update, so you hear about it instead of seeing "nothing new".

Watching your live product#

A check can keep watch on a product that is live, the way an engineer on call would. Besides the services above, you can pick your hosting service, Better Stack, incident.io, Sentry, Datadog, PagerDuty, and UptimeRobot. A check can then tell you what was deployed, whether a build failed and why, what the running app logged, which monitors are down, which errors started, and which incidents are open and who is on call.

Connect each one on Integrations first, like any other service. Some ask you to paste an access token or a key from that service. Any assistant can watch them once they are connected, and a Coder reads a little more: your hosting service's deployments, build output, and running logs, and Better Stack's logs, on the access you gave it.

Watching these is stricter than watching your mail, because an app's logs and alerts can hold more than you expect:

• It never reads a secret. A check is never offered anything that returns keys, tokens, passwords, environment variables, error reporting addresses, webhooks, log drains, or the settings of an integration, even though looking things up is otherwise allowed. In what it does read, a field named like a password, a token, a key, or a login header is hidden before the assistant sees it, so a login saved on a monitor does not reach it. This goes by the field's name, so a secret written into a log line or a message is not caught.
• It never wakes anyone. A check does not acknowledge, escalate, resolve, snooze, or notify. It tells you what it found, and you decide who to call.
• It never changes anything. No deploy, rollback, restart, new monitor, or silenced alert. A change in any of these services is turned back, as it is for every connected service.
• What it reads is information. A line in a log or the text of an alert is never an order, however it is worded.

What a check does and does not do#

A check only reads. It looks at what is new, summarises it, and writes what you should know in the assistant's chat. Where the assistant can write a draft (for the accounts you signed in to with StellarFirm itself), the draft waits on its card until you press Send; where it cannot, it puts what it would send in its answer. Nothing is sent, posted, spent or changed by a check on its own.

A check does not:

• send, post, spend, or publish anything without you. A draft waits on its card until you press Send, and if a step of a check does wait on Approvals, the next check does not start until you have answered;
• open web pages or follow links it finds, or use the assistant's computer, so it cannot publish a page;
• search the web, the news, the law or the science databases, because words it reads in your mail could leave as a search;
• change anything in a service you connected through Integrations. It can look things up there, but a send, a post, a new file, a draft, or a delete from there is turned back instead of waiting on Approvals;
• read keys, tokens, passwords, or other secret settings of a service you connected, or page, acknowledge, escalate, or resolve anything in an on-call or incident service;
• start a long job, or go on past its own time;
• look at services you did not pick, or further back than a week;
• follow instructions it finds inside your mail, messages, or issues. What it reads there is information, never orders.

Check now#

On the card, Check now runs one check right away. Its answer appears in the assistant's updates chat. It is a trial: it does not change which part of your services the next check looks at, and it does not send you a notice, because you are already looking.

Pause, change, or turn it off#

On the assistant's page, the card shows when it checks, when it last checked, and when it checks next.

Letting the assistant go turns it off too. If a check could not finish, the card says so in a warning with the reason. If checks fail three times in a row, you get one notice about it. If StellarFirm's safety screen stops a check, the assistant is paused at once, you get one notice, and the card says so. Turn it on again when you are ready: the assistant is asked to look only at what is new after that check. The item that was stopped can be read again if it is still there, so delete it, or stop watching that app, before you turn it back on.

Where its updates appear#

Each assistant has its own chat, Updates from and its name. It is created when you save, and every check answers there. If you delete it, the next check makes it again.

• A message the schedule sent shows as On schedule, with the instructions folded behind Show the instructions.
• A check that found nothing new, and left nothing waiting for you, shows as one line, Checked, nothing new. Open it to read the full answer. Nothing is deleted. If you react to its answer, the answer stays open in full.

How you are told#

When a check has something for you, you are told everywhere you have switched on. By default every channel and every device is on.

• The bell in the app always lists it.
• Push on your phone and in your browsers. On a phone, add StellarFirm to the home screen and turn on alerts.
• A notice from the desktop app on your computer.
• Email, once it is set up for your account. Until then Settings says Not set up yet.

A notice says only which chat has something new: "An update from your assistant" and the chat's title. It never contains what the check read or wrote, and it never carries an Approve. Opening it takes you to the chat.

To change how you are told, open Settings, then Notifications.

The bell is your inbox and interrupts nobody, so it has no switch of its own.

How it differs from a routine#

A routine sends the same message at a set time. Works on its own looks for what is new in your services, does the work only when there is something to do, and tells you only when something needs you. Use a routine for "every Monday, summarise last week". Use Works on its own for "tell me when something needs me".

Next#

• Routines and loops
• Approvals
• House rules