Automatic tasks
Schedule jobs that run by themselves, in their own chat, and report back to you.
An automatic task runs on its own on its schedule, in its own chat, and reports the result to its owner. Each Nobreinier can have its own.
Create an automatic task
- In the sidebar, open «More» and press «Automatic tasks» (or pin it to the menu with the pin on its row).
- Press «New automatic task».
- Choose the «Owner of the automatic task»: Breini or a Nobreinier.
- Write the instruction: what it must do, verify and report.
- Choose the «Schedule». A date and time written with no time zone, like 2027-03-01T09:00, is read in the automatic task's own time zone.
- Optional: «Maximum runs» (or «No limit»).
- Optional: «Stop when (optional)»: say in your own words when it is done, for example «The invoice is paid». Each run checks it; the run that finds it true switches the automatic task off and tells you why in its chat.
- Optional: «Last day (optional)»: the last day it may run. Its last run is the last one on or before that day; then it ends.
- Save.
When an agent creates an automatic task for you with its checks, a saved flow or a goal, it is created whole or not at all: if one of those parts does not hold, the task is not left half made and the agent tells you what failed.
Start it when something happens
Instead of a schedule, an automatic task can wait for something to happen in an app you already use.
- In «New automatic task», open «When does it start?» and pick the app: «A message arrives in Slack», «Something happens in Linear», «A new error in Sentry» or «An incident in PagerDuty». «On a schedule» is the usual way.
- Fill what the app asks: for Slack, the channel; for Sentry, your organization (or organization/project) and, if you like, search words such as «level:error»; for Linear, a team and search words if you want one; for PagerDuty, a service if you want only one. «Which events» chooses what starts it: for example a new issue, an issue that changes, or an incident that starts, is taken by someone, or is resolved. «Only if it contains (optional)» keeps just the ones with those words.
- Save. Instead of its schedule, the task shows what it waits for, starting with the word «When».
Three more starts need no app. Pick them in the same list:
- «A mail arrives»: it reads the Gmail signed in for this task, and starts only for mail that arrives from now on. Type words to narrow it, like from:laura, or leave it empty for any new mail.
- «A file appears in a folder»: type the folder's full path (like ~/Downloads) and, if you like, the name to look for (like *.pdf). It looks at the folder every few seconds and starts only for a file that appears from now on.
- «A web page changes»: type the page's address (it must start with https://). It looks at the page every few minutes and starts when what it says changes. The address cannot carry a user name or a password.
If an address is not written right, or the folder is not on this Mac or this Mac does not let the app read it, the form tells you in plain words when you press to create the task, and nothing is created.
Under the start, a task you are making now also asks «Tell me only if it finds something» and «Where do I tell you?». The first is the same choice as the funnel on the task's row, so the two always agree. The second is either only the task's chat, or the chat plus a notice from the Mac when it ends. When you choose to be told only if it finds something, the notice comes from the Mac and the question has no other answer.
How it works, in plain words:
- Nothing is opened on your Mac to the outside. The app looks by itself, using the connection you already made: every 30 seconds for Slack, every minute for Linear and PagerDuty, every two minutes for Sentry. No model is called while it only looks.
- The first look only learns what is already there. The task starts only for what happens after you save it, and once for each thing.
- Linear, Sentry and PagerDuty are read with your connection to them (Connections). If PagerDuty is not there, connect it: you paste your PagerDuty key and nothing else. Slack is read with the Slack app of Channels, which must be in that channel; a Slack app made only for the chat (or before this) may need «channels:history» and «channels:read» added in Slack («groups:history» and «groups:read» for private channels), and to be installed again. An app that is not connected shows «connect it first» in the list, and a task whose app cannot be read says why on its row and tries again by itself.
- What arrives from those apps is treated as information, never as orders to the task.
- The same works by asking Breini in plain words: «when a PagerDuty incident starts, tell me what it is about».
A Nobreinier's automatic tasks
- Open the Nobreinier and go to «The Nobreinier’s automatic tasks».
- Press «New automatic task».
- The automatic task belongs to the Nobreinier and uses its account, project and its own browser.
Manage them
- Its list of last runs shows, for each one, its number, when it ran, one plain line of what it found and its cost (when the account bills per use). That way you see which automatic task costs the most.
- Each run says how it ended: «found something», «nothing new» or «a source did not answer». «Nothing new» only shows when every source it read answered. If a site, a connection or a network did not answer (no internet, the sign-in expired, the page changed, the credits ran out), the run says so with the source's name, tells you in its chat and leaves a card in «For you». «Seen» puts it away; when a later run ends without failures, the card goes by itself.
- «Tell me only if it finds something»: the button with the funnel on a task's row, or the same choice in the form when you make the task, or say it when you make the task («tell me only if the price goes up more than 5 %»). Each run says by itself, when it ends, whether it found something. If it found nothing, the task says nothing: no line in its chat, no notice. The run is still there: its list shows «Checked: nothing new» with the time, and its row shows when it last looked and what it last found. If it found something, the notice says what changed in one line («price: 100 → 110», when the run gave both values). A run that fails, cannot tell, or could not read a source is never quiet. «Tell me every time it ends» (the bell) is the other choice; a task made before keeps what it did. Only you turn the quiet choice on, with your own click or your own words: an agent can only make a task speak more, never less. If you change the task's instructions, it forgets what the old ones found, so the next run does not compare against them. A task that runs a saved flow, or that comes from your calendar, has no such choice.
- An automatic task that stopped itself says so on its row, with the reason; one that reached its last day, its last run or its only run says so after its report. Give it a later last day, or turn it back on, if you want it to go on.
- A task that runs often (every 30 minutes or less) can look first, at no cost, at what moves it: new mail, a page, a folder, a connected server, the Clients or the Task Board. If nothing changed since its last run, that turn is skipped and calls no model; its list shows it («N checks found nothing new · no model was called») with what it saved. If a check cannot tell (it could not be read, or was not looked at yet), the task runs as always. A run asked with «Run now» is never skipped. Ask the agent in words: «check my mail every 15 minutes, but only if something came in». A task that repeats a check of web pages also learns this by itself: when a run ended with «nothing new» after reading only pages by their address, the app notes those pages and from then on looks at them first; it calls the model only if they say something different (a clock, an ad or a time like “5 minutes ago” does not count as a change). So every run with nothing new costs less.
- «Run now» starts it without waiting for the schedule.
- «Pause» stops it until you resume it.
- «Waits for cheaper hours · runs at …» on a task's row means its model costs less at that time and the task has no fixed time, so it waits (four hours at most). «Run now» runs it at once. See Cost: price steps and cheaper hours.
- If a run fails, the task tries again by itself, each time after a longer wait (30 seconds, 1 minute, 5 minutes, 15 minutes, then 1 hour). Its chat says how many failures in a row it has had; if you know what it needs, say so there and its next run uses it.
- A task with a daily spending limit does not spend past it, whether the clock or something that arrives starts it: once it reaches its limit, its runs of the day are skipped (or what arrives is held back), counted, and a card in «For you» says what it spent, its limit and how many were skipped. It goes on the next day, or at once if you raise its limit. «Run now» still runs it, because you asked.
- After 10 failures in a row it pauses itself, so it stops spending, and tells you. If what fails is its sign-in, key or permission, it pauses at the first one. Fix that, then turn it back on.
- A Nobreinier that reached the spending limit you set for it (with «Stop its work at the limit» on) does not run its automatic tasks past it: the request is refused before it goes out. That is not a failure: it does not count among the failures in a row and nothing is retried by itself. Press «Go on» in its chat, or raise its limit in «Settings» › «Accounts», and it goes on at once. See Connect your model accounts.
- To pause all of them at once: «Settings» › «Agents & helpers» › «Pause all automatic work».
- If your Mac is asleep when an automatic task is due, it runs when the Mac wakes, as it always did. To have the Mac woken a minute before instead, turn on «Wake my Mac before scheduled work» (see Notifications and startup): only tasks you approved count, and macOS does not wake a Mac with the lid closed on battery.
A change an agent asks for waits for your yes
If a Nobreinier (or Breini) wants to change an automatic task you already approved, the task goes on as you approved it and the change waits for your yes in «For you». That covers the instruction, the schedule, «Maximum runs», «Stop when (optional)», the «Last day (optional)», what wakes it, its checks beforehand, the cheaper hours and the model. A run the agent asks for with words of its own waits too. While you wait, nothing changes: the app says «This waits for your yes in «For you»; nothing changes until you say it.»
- In the sidebar, press «For you».
- Find the «Automatic task» card: it says who asks to change the task and what they want to change.
- Press «Yes, do it» to make the change count, or «Reject» to leave the task as it was.
Your yes counts for the card you saw. If the task changed after that, the app says «This task changed after you saw it. Look at it again before you answer.» and applies nothing until you answer the new card.
What you ask Breini in your own words in the chat is applied at once, with no card. And a task that only moves its own next run, no more often than you approved and changing nothing else, does not ask for your yes either.
A task that paused itself waits for you in «For you»
A task pauses itself when it fails many times in a row, or when its sign-in, key or permission stops working. Because that can happen while you are not looking, the app does not settle for a notice that goes away: it leaves a card in «For you», tagged «Paused», that says which task paused itself and why, in plain words (for example, that it failed 10 times in a row and paused itself so it stops spending). It has two buttons:
- «See the task» takes you to «Automatic tasks» with that task chosen, to see its history and what failed.
- «Resume» switches it on again: from then on it runs on its schedule.
The card goes away by itself once the task is on again. If what failed was its sign-in, key or permission, fix that before pressing «Resume»; otherwise it will pause itself again.
Automatic flows: steps that do not spend a model
A flow is a list of steps that runs by itself, one after another, without asking the model for anything while it does not need to. It is for a job that repeats the same way: look whether something new arrived, prepare the data and, only if there is something, tell you or send it. If nothing is new, the flow ends right there and costs nothing.
There is no button to build one: you ask a Nobreinier in your own words («every hour look for new invoices and tell me») and it saves the flow. A flow that runs on a schedule is started by an automatic task, and putting a flow on its schedule (or taking it off) is your decision.
The steps you will see
While a flow works, it shows in the chat, in a box above the message box, with how many steps it has done and how much it has spent. Press «Show the steps» to open it. Each step says what it is:
| Step | What it does |
|---|---|
| «Tool» | Uses a tool of the app or a recipe, with the same permissions and notices as always. |
| «Anything new?» | Compares with what it saw last time. If nothing is new, the flow ends right there, at no cost. |
| «If» | Picks one way or another according to what it found. |
| «Prepare data» | Sorts, filters, counts or joins what was read. |
| «Join» | Brings the ways together and goes on with the one that came. |
| «For each» | Repeats its steps for each item of a list. |
| «Agent» | One single call to the model, which reads only its instruction and the data it is given, never your conversation. |
The other steps are «Command», «Wait» and «Approval» (the flow stops and asks you). Each step shows its state: «waits», «ready», «working», «done», «failed», «skipped» or «not on this way» (the way that was not taken). A step that fails can go on, try again or leave by its error way.
A flow can also have a step that works on one piece of work until it is good: the AI does it, has it reviewed against checks you approved («Check: …», one per line on the card), and redoes only what was missing. It stops when every check passes. If a round changes nothing, or leaves the same checks unmet as an earlier round, it stops there and the step shows as failed with the best result so far, never as done. Because there is no fixed number of rounds, a flow with this step must have a budget before you can approve it, and every round counts against it. If your request in a round differs from a rule you saved, your request wins for that round, the saved rule stays as it was, and the AI says so in one short line; safety and permission rules win over both.
What the last time cost
On a saved flow's row you see how it went: «last run: nothing new · … s · no cost» when there was nothing to do, or «last run: … · … s · US$…» with what it spent. «Done», «Did not finish» or «Stopped» say how it ended. «Pause», «Resume», «Stop» and «Retry» work as in any long job; if its budget runs out, «Its budget is spent: raise it to go on» and «Double the budget» put it at twice as much.
When you repeat something, the app proposes the flow
If you, or one of your Nobreiniers, do the same thing three times or more (the same tools in the same order, with different values each time) and one of them was in the last two weeks, the app notices it in the records it already keeps on your Mac and proposes it to you already built in «For you»: «You did this … times: shall I turn it into an automation?», or «… did this … times: shall I turn it into an automation?» if it was a Nobreinier. The card says «It is saved as …», its steps and what changes: how long it took each time and that the AI did it («Each time the AI did it, and it took … s.») against how it would run («As an automation: the AI does not work, about … s.»).
- What stays the same from one time to the next is written into the step; what came out of an earlier step is taken from that step; and where judgement was needed (which click to make, which comment to write) the model works only at that point. What you or a customer asked each time cannot be part of the flow, because the flow could not know it.
- If what it repeats only reads, the flow ends with «Anything new?»: «It looks whether anything is new: if nothing is, it ends here at no cost.»
- Nothing runs and nothing is saved until you approve it. «Review and approve» («Approve these steps: from now on it is done this way») checks it as any flow and saves it; then the app tells whoever repeated it to use it instead of repeating the steps. «No, thanks» («I will not propose it again») remembers it: that procedure is not proposed again.
- Nothing the app reads to propose it leaves your Mac.
Your approval before it runs by itself
A saved flow does not run by itself until you approve it. Meanwhile its row in «Workflows waiting for an event» says «Waits for your approval: it does not run until you approve it».
- Open «For you». The flow is there with the tag «Automation», its steps, one per line and said as each permission card says them (who does what, on which page or address, with no tool names), and, on separate lines, everything it sends out of your Mac: «Leaves your Mac for: …» with the places it goes, each one once, «Uses the saved key …» if it uses a saved key, and «Something that reaches you from outside can decide …» if something that comes from a stranger could decide, for example, who receives it or the web address, and in which part of what leaves. The card also shows what it costs: «Expected cost per run: about …» with the mean of its last measured runs (up to ten), and «A month of it: about …» with the runs a month, from the schedule of the routine that starts it or, if none does, from how often it really ran lately. If its runs are covered by your plan, the card says «Covered by your plan: no price per run» and «A month of it: covered by your plan, no price», with how many runs a month, and never makes up a price. If it has not run yet, it says «Expected cost: not known yet, none of its runs has been measured»; the figures are measured, never a guess.
- Read it. If it is all right, press «Review and approve» («Approve this version: from now on it runs by itself»). You can also press «Approve» on that row.
- From then on it runs by itself with that version. If a Nobreinier changes it, the flow says «Changed since you approved it:» and, below it, one step per line, which is new, which changed and which was removed, and waits for your approval again; the version you approved is not changed behind your back.
What a flow reads from outside (an e-mail, a page) is treated as data, never as orders: it cannot decide where something is sent, unless you approved that when you sealed it. Keys are named, never written inside the flow, and each key works only on its own site. There are tools a flow never uses: the ones that change the app's rules, its keys or its own work.
The flows you already had
Flows saved before this version will not go on running by themselves: when the app opens they are switched off until you approve them. You will see them all in «For you» with «Review and approve»; approve them one by one, after reading what they do and what they send. The ones that wait for an event (an e-mail, a file, a page) are in «Workflows waiting for an event», each with «Stop waiting for this event».
